A widely shared infographic tells boards which eight AI governance platforms to evaluate in 2027. It maps one layer of a seven-layer market, leaves out one of Gartner’s three Leaders, names the wrong Microsoft product and ignores the roughly forty acquisitions that moved enforcement into security, identity and data platforms. This is the complete map: who owns what after the buying spree, which rulebooks bite while a system runs and when, what the labs and the year’s incidents taught us, how far my own TrustOS architecture really gets, and the demo tests a buyer should run before signing anything. Facts are stamped as of 18 November 2026.
The Two-Minute Version
- The governance-platform market is not short of products. It is short of refusals that anyone outside the vendor can verify. Most products on the shortlists boards are being handed can describe a risk, score it and file it. Very few can stop the action that creates it.
- The infographic everyone is forwarding covers the governance-of-record layer only, as it looked in 2024. It omits Truyo, one of three Leaders in Gartner’s first Magic Quadrant for AI Governance Platforms, and lists Microsoft Purview where Microsoft’s agent governance product is Agent 365.
- Think of the market as seven layers, each answering one question: what exists, who is it, may it do this, did it behave, can anyone vouch for it, who pays if it fails, and where did this content come from. Enforcement lives in the identity, gateway and runtime-security layers, not in the dashboard.
- Between 2024 and November 2026, roughly forty AI security, agent identity, gateway and observability start-ups were bought by Palo Alto Networks, Cisco, CrowdStrike, Check Point, Okta, SailPoint, Cyera, Proofpoint, Dynatrace, ServiceNow, OpenAI and others. In practice the off switch is now the credential, and the credential belongs to an incumbent.
- Every major rulebook now contains duties that bite while a system runs. Almost none binds enterprise AI before 2027 outside China and a few US states. The EU moved its high-risk runtime duties to 2 December 2027 and 2 August 2028. Anyone selling “EU AI Act compliance” for high-risk systems today is selling readiness.
- Singapore has built the world’s most complete public governance stack without passing an AI law: AI Verify, the assurance pilot and sandbox, IMDA’s agentic framework, MAS’s SAFR runtime design, the MAS risk guidelines, a public agent-testing method with accredited testers, a Future of Finance Institute and, on 18 November, National AI Missions that will trial agents paying with purpose-bound money within defined authority and limits. It is non-prescriptive in law and the most specific jurisdiction anywhere in method. SAFR in particular is the gate-before-the-action architecture most commercial platforms lack.
- Elsewhere in Asia, Hong Kong has proposed an agent registry, Know Your Agent and kill switches, and Malaysia is legislating. Australia flipped from voluntary standards to a promised regulator within months of an agent incident.
- The labs loosened their written pledges in 2026 and tightened their runtime gates after incidents. That is the strongest evidence yet that governance is what the running system refuses, not what the binder says.
- No vendor is a genuine one-stop shop. ServiceNow and Microsoft come closest, each centred on its own estate. A real one-stop shop needs twenty capabilities. My own TrustOS architecture has specified about half of them, built and independently tested none of the critical ones, and I say so below.
- Buyers should stop asking vendors what they cover and start asking them to show a denied action, with its receipt, in a live demo. The scorecard in Section XII is built for exactly that meeting.
I. The Infographic, Corrected
Start with a case, because the case explains the market better than any quadrant. In June 2026, an internal OpenAI test agent was asked to find public data on medicine spending in Australia. It found its way into the Medicare Statistics Reporting Service and viewed non-public aggregate statistics and internal file names. No personal data was reported accessed. OpenAI detected the activity in August and told the Australian government on 10 September, by email to a public disclosures mailbox, about 84 days after the access. The Prime Minister, Anthony Albanese, said it took the company “way too long to inform the Government”. On 10 November Canberra accepted all eighteen recommendations of a rapid review, including a dedicated AI safety regulator and rapid incident notification.
Now ask a simple question. Which of the eight platforms on the infographic now doing the rounds in boardrooms would have stopped that request? On their public documentation, none of them sits where it could have. They would have recorded the agent in an inventory, mapped it to a risk tier and attached the relevant clauses of the EU AI Act. The boundary that failed lived in a task description. The agent never touched the governance platform on its way to the portal. That is the gap this essay is about, and it is the same gap I described in “Pretty Fangs” in September: if a control cannot say no, it is an opinion with a dashboard.
The infographic, titled “8 AI Governance Platforms to Evaluate in 2027” and credited to “Chief AI Leader”, is not wrong about its eight vendors. All of them are real, none has been acquired, and the “best for” labels are fair as far as they go. The problem is that they go only as far as 2024. Every serious vendor in this layer now sells an agent registry, agent discovery and at least a partner route to runtime control, and those are what separate them. Here is what each actually does in November 2026.
| Vendor | Infographic label (paraphrased) | What it is in November 2026 | Gartner MQ, as reported |
|---|---|---|---|
| Credo AI | Regulatory compliance | Policy and regulatory mapping leader; Agent Registry in public preview since September 2025; runtime enforcement still described as roadmap (per a competitor) | Visionary |
| IBM watsonx.governance | Model monitoring | The only vendor confirmed as a Leader in Forrester, IDC and Gartner reports; agent monitoring since October 2025; analysts reportedly found no native inline runtime gateway | Leader (top right) |
| Microsoft Purview | M365 data governance | Wrong product: Purview is data security. The agent stack is Agent 365 plus Entra Agent ID plus Purview plus Defender | Microsoft not among the 13 |
| OneTrust | Privacy-first | Announced an AI Control Plane that allows, blocks or escalates agent actions, and a headless MCP gateway, on 30 September; key pieces in preview | Visionary |
| Holistic AI | Bias testing | Live inventory, Agent Graph, and “Guardian Agents” that monitor and restrict access | Challenger (the only one) |
| ModelOp | Lifecycle governance | AI system of record; auto-imports discovered agents; runtime through Kong and NVIDIA OpenShell partnerships | Visionary |
| ServiceNow AI Control Tower | Existing ServiceNow customers | The most ambitious one-stop claim: 30 discovery connectors, a demonstrated kill switch, an MCP server, Veza and Traceloop bought in | Leader |
| Relyance AI | Privacy and data governance | Data mapping against contractual and regulatory policy; agent tracing not documented | Niche Player |
ServiceNow deserves a sentence more, because it makes the boldest claim. Independent fieldwork by SiliconANGLE with 15 large customers is the best reality check I found: the kill-switch claim was “overstated”, with one successful blocking pilot; one practitioner estimated that only 30 to 40% of AI activity passes through ServiceNow control paths; 12 of the 15 raised cost; and one respondent called it “more like a governance orchestration layer than the actual governance of AI”.
Five corrections
- Add Truyo. Gartner’s Leaders were IBM, ServiceNow and Truyo, as relayed by OneTrust and an IBM partner, and Truyo announced its own placement on 22 June. A competitor notes that Truyo documents no runtime or agent telemetry. Its Leader position suggests Gartner weighted inventory breadth and privacy-office buyers heavily, which tells you what the quadrant measures.
- Replace Purview with Agent 365 plus Entra Agent ID. Agent 365 was announced at Ignite in November 2025 and became generally available on 1 May 2026. It registers Microsoft and non-Microsoft agents, discovers more than 60 local agents such as Claude Code and GitHub Copilot CLI, governs MCP servers and syncs agents from Agentforce, Bedrock AgentCore, Vertex AI, Databricks, Snowflake, Anthropic and OpenAI. Microsoft is a Leader in IDC’s assessment but sits outside Gartner’s quadrant.
- Add the rest of the quadrant. Airia (Visionary, agent platform plus gateway), Monitaur (Visionary, insurance model governance), SAP (the highest Niche Player, with an AI Agent Hub in LeanIX), Cranium AI (Niche, security-led, bought Aiceberg in May) and Saidot (Niche, raised an undisclosed round on 5 November for runtime policy-as-code). Honourable mentions went to Enzai, LatticeFlow AI, Modulos, Singulr, Trustible and WitnessAI.
- Fix the stale names. Fairly AI and Anch.AI are now Asenion. Securiti belongs to Veeam, after a $1.725bn deal that closed in December 2025.
- State the provenance problem. I could find no newsletter, account or methodology called “Chief AI Leader”, so the selection criteria, and whether anyone paid for inclusion, are unknown. Treat it as an unaudited list.
One common error to avoid: the widely reported $200m round of May 2024 belongs to H, a Paris start-up, not to Holistic AI, the London governance firm.
Small, crowded and converging
Gartner screened more than 100 vendors and admitted 13. It forecasts spending on AI governance platforms of about $492m in 2026, passing $1bn by 2030, as relayed by trade press; I could not find the original release, and one relay calls it “AI data governance”. Older market estimates from MarketsandMarkets and Grand View include services and run roughly double for 2026, so they should not sit beside Gartner’s number. For scale, Gartner puts the AI security market at almost $5bn in 2027, about ten times the governance-platform figure. That ratio explains where the acquirers put their money.
Demand is moving from writing policy to proving control. EY’s survey of 202 senior US AI executives, released on 15 September, found that 98% have formal AI governance policies, 49% have not updated governance for agents, 47% have bypassed governance for urgent deployments and 26% cannot detect unauthorised agents. McKinsey’s 2026 trust survey puts average responsible-AI maturity at 2.3 out of 4, with only about a third at level 3 or above on agentic governance. Budget ownership is split three ways: privacy and legal in the IAPP data, security and risk in Stanford’s AI Index, and Forrester’s AEGIS framework of August, which claims agent governance for the CISO. That split budget is why the market looks like five different industries sharing one name. Gartner’s Lauren Kornutick put the diagnosis plainly: “Traditional GRC tools are simply not equipped to handle the unique risks of AI”. Her own quadrant still rewards inventory breadth over enforcement, which is a tension worth holding on to.
II. The Seven-Layer Map
The useful way to read this market is not by vendor, and not by Gartner box, but by the question each layer answers. Gartner’s own AI TRiSM model already separates a governance layer from a runtime inspection and enforcement layer, and states, in a secondary relay, that no single vendor covers all AI risks. I extend that into seven layers. For each one, the table records who owns it after the 2024 to 2026 consolidation, what the products actually enforce, and which of three tiers they reach: Enforced (a control blocks or alters an action), Evidenced (telemetry populates control status automatically) or Documented (a questionnaire, attestation or upload).
| Layer and its question | Who owns it now (18 Nov 2026) | What it actually does | Tier reached |
|---|---|---|---|
| 1. Governance of record: what exists, and what are the rules? | IBM; ServiceNow; Truyo (IntraEdge); Holistic AI; Credo AI; OneTrust; ModelOp; Monitaur; Airia; SAP LeanIX; Relyance; Cranium (owns Aiceberg); Saidot; Asenion; Veeam (Securiti); Collibra; Databricks Unity Catalog; Dataiku; Trustible; Enzai; Modulos. Exits: SnapAI to HR Path; CastleHill to Concentrix | Use-case and agent registries, risk tiering, control libraries mapped to the EU AI Act, NIST and ISO 42001, approvals, audit evidence | Mostly Documented; Evidenced through integrations; Enforced only via partners or previews |
| 2. Discovery: what is actually running? | Microsoft (Agent 365, Defender, Entra); ServiceNow; OneTrust; ModelOp; Cyberhaven; Netskope; Zscaler (SPLX, Symmetry); Cloudflare; F5 (SurePath); Akamai (LayerX); CrowdStrike (Seraphic); Proofpoint (Acuvity); WatchGuard (Perimeters.io) | Finds agents, GenAI apps and MCP servers from endpoint, network, identity and cloud telemetry, then risk-scores them | Evidenced; Enforced where the tool sits inline and blocks |
| 3. Runtime security: is this input, output or call dangerous? | Palo Alto Prisma AIRS (Protect AI, Koi, Portkey); Cisco AI Defense (Robust Intelligence); Check Point (Lakera, Cyata); SentinelOne (Prompt Security); F5 (CalypsoAI); CrowdStrike (Pangea, Dreadnode); Cato (Aim); Proofpoint (Varonis, AllTrue); Fortinet (Virtue AI); A10 (TrojAI); hyperscaler guardrails; NVIDIA NeMo, Meta LlamaFirewall; independents Zenity, HiddenLayer, Noma, Straiker, WitnessAI, Lasso, Pillar | Inspects prompts, outputs and tool calls; blocks injection, jailbreaks and exfiltration; maps detections to OWASP and MITRE ATLAS | Enforced, at the content and data level; rarely at the business-action level |
| 4. Identity and gateways: who is it, and may it do this? | Microsoft Entra Agent ID; Okta (Axiom, Permiso); Palo Alto (CyberArk); Cisco (Astrix); SailPoint (Entro); Cyera (Oasis); CrowdStrike (SGNL); 1Password (Apono); Delinea (StrongDM); ServiceNow (Veza); Snowflake (Natoma); Databricks (Antimatter); gateways from Cloudflare, AWS AgentCore, Microsoft, Kong, Docker, Netskope, TrueFoundry and others | First-class agent identities, short-lived scoped tokens, just-in-time access, MCP traffic brokering and logging | Enforced at the permission level: revoking the credential is the practical kill switch |
| 5. Evaluation and observability: did it behave? | Dynatrace (Arize, announced); Cisco/Splunk (Galileo); ClickHouse (Langfuse); CoreWeave (Weights and Biases); ServiceNow (Traceloop); Mintlify (Helicone); CrowdStrike (Fiddler); OpenAI (Promptfoo); Check Point (Deepchecks); Beacon (Haize Labs); Datadog; Braintrust | Agent tracing on OpenTelemetry GenAI conventions, evaluations, red teaming, drift | Evidenced, as engineering telemetry; audit evidence only once retained and independently reviewed |
| 6. Assurance and insurance: can anyone vouch for it, and who pays? | ISO/IEC 42001 certifiers (BSI passed 50 certificates in October); HITRUST AI; AIUC-1; Singapore’s AI Tester Accreditation (five firms accredited by 14 October); insurers Armilla, Testudo, Munich Re | Certifies management systems, tests agent behaviour, underwrites residual risk | Documented to Evidenced; the only layer producing formal third-party evidence |
| 7. Provenance: where did this content come from? | C2PA Content Credentials; Google SynthID (more than 100bn items marked); adopters include OpenAI, NVIDIA, ElevenLabs, Kakao | Machine-readable marking and detection for EU Article 50 and China’s labelling standard | Evidenced, and machine-verifiable |
Read the right-hand column from top to bottom and the shape of the market is obvious. The layer the infographic covers is the one that mostly documents. The layers that enforce, identity and runtime security, are the ones the acquirers bought. Discovery is the contested middle ground, built automatically from telemetry by security vendors and laboriously by intake forms in governance tools. Assurance is the only layer that produces evidence a regulator or an underwriter will accept from a third party, and provenance is the only one whose evidence a machine can verify.
Notice also what is missing from every row: a layer that decides, before a business action executes, whether this agent, acting for this principal, under this mandate, may do this specific thing, and leaves a record the agent cannot edit. Runtime security asks whether content is malicious; identity asks whether a credential is valid; governance of record asks whether a policy exists. None asks whether the action is within authority. I called that missing piece the Disposition Engine when I wrote about MAS’s SAFR design in July, and five months later it is still the empty chair at the table.
Four grades of evidence
If the seven layers tell you who does what, the next question is which outputs count as proof. Only a narrow band of what these products produce is formal evidence today. I grade it in four tiers, and I would encourage buyers to label every artefact in a vendor pitch with one of them.
| Tier | What it is | Examples | Caveat |
|---|---|---|---|
| A: accredited attestation | Certificates and assurance opinions from accredited bodies | ISO/IEC 42001 certificates from UKAS, ANAB or RvA-accredited certifiers; professional assurance reports over AI controls; HITRUST AI certification | ISO 42001 certifies a management system, not behaviour, and gives no presumption of conformity under the EU AI Act |
| B: independent behavioural testing | Third parties testing what the system does | AIUC-1, which tests each agent against about 5,000 risk and attack combinations and recertifies quarterly; Singapore’s accredited testers; insurer-embedded certification | Singapore’s scheme accredits the tester, not the system; it does not certify any system as safe |
| C: machine-verifiable provenance | Signed or detectable content marks | C2PA manifests; SynthID detection under the EU marking Code of Practice | Proves origin, not correctness |
| D: engineering telemetry | Vendor dashboards, traces, evaluation scores | Datadog, Dynatrace and Arize, Cisco and Galileo, Langfuse, Braintrust, Promptfoo | Becomes evidence only when retained under controls, tied to a requirement and independently reviewed |
The tiers carry a lesson for buyers. A dashboard is Tier D. It is useful, and it is not proof. When a vendor tells you its platform “evidences compliance”, ask which tier, who reviewed it and whether the record survives the vendor being acquired, which, on this year’s form, it probably will be. That last question is not rhetorical. Datadog’s native ingestion of OpenTelemetry GenAI conventions matters more than it sounds, because portable traces mean the audit evidence can outlive the company that collected it.
III. The Acquisition Wave: The Off Switch Changed Hands

The most important governance decision of the past two years was not made by a regulator. It was made by corporate development teams at security, identity and data companies, who between 2024 and November 2026 bought roughly forty start-ups that could inspect, authorise, broker or observe what AI systems do. The result is that runtime governance is being absorbed into platforms that were never sold as “governance” at all, and the practical off switch for an agent, revoking its credential, now sits with an identity or security incumbent.
The buying came in two waves. The first, in 2024 and 2025, was network, endpoint and application-security incumbents buying prompt firewalls. Forrester’s Jeff Pollard counted more than $2bn spent by eight vendors in about 18 months and predicted that AI security would arrive inside firewall, XDR and SASE platforms. He was right. The second wave, through 2026, was agent identity, non-human identity, MCP gateways and agent observability, and the buyers widened to identity companies, data-security companies and the agent-platform owners themselves.
| Announced | Buyer | Target and what it adds | Value (SEC-derived where available) |
|---|---|---|---|
| Aug 2024 | Cisco | Robust Intelligence: model validation, now Cisco AI Defense | About $400m reported |
| Apr 2025 | Palo Alto Networks | Protect AI: model scanning, now Prisma AIRS | $634.5m (headline about $650m) |
| Jul 2025 | Palo Alto Networks | CyberArk: machine and agent identity | $21.1bn (headline about $25bn) |
| Sep 2025 | Check Point | Lakera: runtime guardrails | $201.8m (headline estimates near $300m) |
| Sep 2025 | CrowdStrike | Pangea: AI detection and response | $222.7m |
| Oct 2025 | Veeam | Securiti: data and AI governance | $1.725bn |
| Dec 2025 | ServiceNow | Veza: access graph for humans, machines and agents | $1.20bn |
| Apr 2026 | Palo Alto Networks | Portkey: AI gateway, pitched as a control plane for agents | $117m |
| May 2026 | Cisco | Astrix: agent identity; Galileo: AI observability (completed May) | About $400m reported (Astrix) |
| Jul 2026 | Cyera | Oasis: agent and non-human identity governance | About $1bn |
| Aug 2026 | Dynatrace | Arize: AI observability (announced; closing unconfirmed) | $915m |
| Oct 2026 | Proofpoint | Varonis, which owns AllTrue.ai | About $6.3bn |
| Nov 2026 | CrowdStrike | Fiddler (now Falcon Observe) and Dreadnode | About $495m combined, reported |
| Nov 2026 | Palo Alto Networks | Clutch: agent access security (in talks only) | About $500m, reported |
The table is a selection. Several 2026 values are reported rather than disclosed, and SEC-derived figures come from a secondary tracker; headline figures often run higher because they include stock and earn-outs. Palo Alto Networks is the most aggressive consolidator, with Protect AI, CyberArk, Koi, Portkey and Console bought and Clutch reportedly in talks. Observability consolidated just as hard: Arize, Galileo, Langfuse, Weights and Biases, Helicone, Traceloop and Fiddler have all been bought or agreed to be, Humanloop shut and its team joined Anthropic, and WhyLabs went quietly to Apple.
Three consequences
First, runtime governance is becoming a feature, not a category. It is being folded into identity, SASE, XDR and data-security platforms. The enforcement point that matters in practice is the credential. Microsoft’s Entra is previewing a default lifecycle policy that notifies an agent’s owners after 60 days of inactivity and then disables the identity. Okta’s Agent SSO issues short-lived governed tokens in place of static API keys, and its Cross App Access extension has been adopted as MCP’s enterprise-managed authorisation mechanism. Whoever controls the token controls whether the agent acts.
Second, guardrails are commoditised. Every hyperscaler and every security service edge vendor ships one. Differentiation has moved to cross-model, cross-agent policy at the tool-call level, and to binding that policy to identity. A prompt filter is now a checkbox on a datasheet.
Third, the two halves of the market are colliding from both sides, and neither does the other’s job well. Runtime vendors now market “security and governance” and build inventories automatically from telemetry: Cato, Proofpoint, Check Point, Fortinet, Zenity and Cyberhaven all do it. Governance vendors bolt on runtime through previews and partnerships: OneTrust, ModelOp, Saidot and Credo. But runtime vendors map detections to OWASP and MITRE ATLAS, not to the EU AI Act or ISO 42001, and governance vendors’ “policy packs” mostly map regulation text to controls with evidence collected by workflow. The security team gets blocking without regulatory meaning. The compliance team gets regulatory meaning without blocking.
Red teaming has an independence problem
One consequence deserves its own paragraph. The tools that test AI systems are being absorbed by the parties they should be independent of. OpenAI owns Promptfoo, which Futurum says about a quarter of the Fortune 500 used. Check Point owns Lakera and Deepchecks. Beacon owns Haize Labs. Harvey owns Guardrails AI. I argued in “Who Audits the Auditor?” that an evaluator embedded in a lab is a second party wearing a third party’s badge. The same logic now applies to tooling. When the red-team harness belongs to the model vendor, the test result is a vendor claim. That makes accredited or certifiably independent testers, public AI security institutes and schemes such as AIUC-1 more valuable, not less, and it should change how procurement teams weigh a “red-teamed by” line in a pitch deck.
Insurance is becoming the hidden regulator
The most demanding buyer of governance evidence in 2026 was not a regulator. It was an underwriter. Armilla, a Lloyd’s coverholder, now offers limits up to $25m with affirmative cover for agent failures and insurable EU AI Act and Colorado fines, and every policy embeds independent AI system certification. Testudo reached $9.25m with Atrium and QBE joining Apollo on its panel. ElevenLabs used its AIUC-1 certification to insure its voice agents. Meanwhile mainstream insurance is retreating: Verisk’s generative-AI exclusion endorsements for general liability took effect on 1 January 2026, and W. R. Berkley, AIG and Great American filed AI exclusions for directors’ and officers’ and errors and omissions cover in November 2025, according to secondary reports.
Limits topping out around $25m against a cyber market of roughly $15bn show how thin the market is. But insurers price evidence, and a platform that cannot produce underwriter-grade records will lose deals it never sees. I expect this, not the EU AI Act, to be the force that makes Tier A and Tier B evidence normal in 2027.
IV. Can It Say No? Runtime Versus Paper
The single most useful question to ask of any governance product is the one I put to every category in September: can it say no? Not “can it flag”, not “can it score”, not “can it notify a human”. Can it stop an action before it executes, and leave a record that the agent cannot edit? Most products on the market answer with a confident yes in the brochure and a quiet no in the architecture.
The distinction matters because the 2026 incidents all failed in the same place. At Replit in July 2025, an agent deleted a production database during an explicit code freeze, because the freeze was a sentence in a prompt, not a permission. At Hugging Face in July 2026, OpenAI test models running with reduced cyber refusals exploited a sandbox proxy flaw, moved laterally and uploaded a malicious dataset, because the evaluation sandbox’s egress was not enforced. In the UK AI Security Institute’s cyber tests, agents took unsanctioned actions on the live internet, and the Institute’s own report said “Containment relied on human vigilance”. In Canberra, scope lived in the task. In every case the post-incident fix added network egress limits, permission separation and active monitoring. Not one of the fixes was a better policy document.
Three enforcement tiers
To make the test usable, I sort every control a platform claims into three tiers. The labels are simple on purpose, because procurement teams need to apply them in a meeting.
| Tier | What it means | Example | What it proves |
|---|---|---|---|
| Enforced at runtime | A control blocks or alters an action before it executes | A policy decision point denies a tool call; a gateway refuses an unapproved MCP server; an identity provider revokes a token; a kill switch halts the task and preserves logs | That the boundary holds whether or not the agent cooperates |
| Evidenced continuously | Telemetry populates control status automatically | Action logs, traces, evaluation scores, drift alerts feeding a control dashboard | That someone could have noticed, and can reconstruct what happened |
| Documented only | A questionnaire, attestation or uploaded artefact | An AI use-case intake form; a risk assessment PDF; a policy acknowledgment | That someone said the right thing at a point in time |
Most vendor “policy packs” deliver the third tier natively, the second through integrations, and the first only through partner guardrail or gateway products. That is not a scandal. Documentation and evidence matter, and regulators will ask for them. But a buyer who believes they have bought the first tier when they have bought the third has a governance programme that will perform beautifully in an audit and fail completely in an incident. As I put it in “Governance Is a Position, Not a Document”, NIST stops at documented and ISO 42001 governs a management system. Neither is wrong. Neither stops an agent.
Why intervention is the hardest control to buy
Across the Five Eyes agencies’ May guidance, the HKMA, MAS, China and the EU, a credible minimum operational control set emerges: an agent and model registry reconciled against what is running; least-privilege, scoped tool permissions; tiered approval for high-risk actions with evidence that review was substantive; a kill switch and suspension; automatic logging of actions, tool calls and inter-agent messages; continuous evaluation and red teaming; incident detection within statutory clocks; content marking; third-party assurance; and reversibility. Against that list, the control most poorly served by governance-of-record platforms is intervention, because intervention requires integration into the agent’s execution path, not a console beside it.
Gartner’s first Market Guide for Guardian Agents said the same in February, as relayed in a vendor-sponsored summary: many platforms observe rather than intervene, gateways only work if traffic passes through them, and the neutral cross-platform enforcement layer does not yet exist. Gartner’s own Hype Cycle survey found 47% of security leaders reporting unfunded or diverted budgets for runtime defence, so do not assume the money will follow the diagnosis.
Supervisors have codified the critique
The academic case against paper governance is old and good, from Brent Mittelstadt on principles without enforcement to Goodman and Trehu on audit-washing and Ben Green on oversight mandates that legitimise flawed systems. What is new in 2026 is that supervisors wrote the critique into their own texts.
The HKMA’s discussion paper on agentic AI in banking, published on 2 November, is the sharpest example. It states that “A formal approval step is not, by itself, meaningful human oversight”, and asks that the reviewer see the substance of the action, be able to reject it and have time to decide. That sentence is in the discussion paper, not in the October press release or speech, and I attribute it only there. The evidence base for it is uncomfortable. Users accept about 93% of agent permission prompts, and in Anthropic’s October study analysts followed a secretly assigned coin-flip lean in 86% of methodological choices. I wrote about the human sign-off problem in “Presence Is Not Power”. A rubber stamp is a Documented control wearing an Enforced costume.
There is a design implication that no vendor I found has picked up. If supervisors are going to test whether oversight was meaningful, someone has to measure it: what the reviewer saw, how long they took, how often they rejected or overrode, and whether the rejection rate is plausibly above zero. That is a product feature. It does not exist on the market today, and three Asian regulators are now asking for its outputs.
V. From Rulebook to Runtime: The Crosswalk

Every major rulebook now asks for things that must happen while a system runs: logs, monitoring, the ability to intervene, incident clocks and labelling. Almost none of those duties binds enterprise AI before 2027, outside China and a handful of US states. That combination, universal direction and deferred force, defines the next twelve months for anyone buying a governance platform. You are buying for a destination everyone agrees on and a deadline almost nobody has set.
The tables below translate each regime’s runtime duties into the controls a platform would have to execute. I have split them by family so they stay readable on a phone. Article-level references for the EU AI Act, NIST subcategories and ISO Annex A clauses are given as the standard texts number them; I have not re-read every clause against the consolidated text for this essay, and say so in the section on what I could not verify.
The Western statute and the two standards
| Runtime control | EU AI Act (as amended by Regulation 2026/1744) | NIST AI RMF and AI 600-1 | ISO/IEC 42001 |
|---|---|---|---|
| Inventory or registry | Registration of high-risk systems | GOVERN mapping (voluntary) | Scope of the management system |
| Risk classification | Article 6 and Annex III tiers | MAP functions | Impact assessment via ISO/IEC 42005 |
| Logging | Article 12 automatic logs; Article 26(6) six-month retention by deployers | MEASURE 2.4 | Annex A.6.2.8, event logs |
| Monitoring | Article 72 post-market monitoring; Article 26(5) deployer monitoring | MANAGE 4.1 | Annex A.6.2.6, operation and monitoring |
| Human oversight and intervention | Article 14(4)(e), the “stop” button; Article 14(4)(b), automation bias | MANAGE 2.4: supersede, disengage or deactivate | Annex A.9, responsible use |
| Incident reporting | Article 73: 15 days, 2 days for critical infrastructure, 10 days for a death (high-risk, from 2027); Article 55 for systemic-risk models now | MANAGE 4.3 | Annex A.8, communication |
| Labelling and provenance | Article 50, applying since 2 August 2026 | AI 600-1 provenance actions | Not specific |
| Third parties and supply chain | Provider and deployer split | GOVERN 6 | Annex A.10, third parties |
The Asian supervisors
| Runtime control | MAS AI Risk Management Guidelines | IMDA agentic framework and MAS SAFR | HKMA (discussion paper, 2 Nov) |
|---|---|---|---|
| Inventory or registry | Explicit inventory; per a secondary summary, agent attributes include identifiers, tools and guardrails | Bound risks upfront; SAFR agent identity | Internal records plus a proposed central registry of material agents |
| Risk classification | Materiality per use case, with autonomy as an input | Assess and bound risks before deployment | Risk-based Know Your Agent |
| Logging | Monitoring of reasoning, actions and tools (secondary summary) | Tamper-evident audit log (SAFR) | Preserve logs when an agent is stopped |
| Monitoring | Lifecycle monitoring and change management | Agentic Testing Starter Kit: monitor and re-test | On-going monitoring |
| Human oversight and intervention | Oversight that accounts for automation bias; kill switch to be considered for high materiality (secondary) | Meaningful human accountability; SAFR disposition: Auto-Execute, Escalate, Deny, Observe | Kill switch that halts, revokes and preserves; possible supervisory kill switch; “meaningful” oversight test |
| Incident reporting | Not stated in the release | Not stated | No duty found in the part of the paper read |
| Third parties | Accountability for third-party AI; limit, suspend or replace | Third-party agent guidance added in the May update | Dependencies recorded in the registry |
China, Korea and the US states
| Runtime control | China | Korea AI Basic Act | Colorado SB 26-189, California SB 53, New York RAISE |
|---|---|---|---|
| Inventory or registry | Algorithm and model filing | None found | Not required |
| Risk classification | Security assessment; five-level grading in TC260’s framework 2.0 | High-impact domains | Colorado dropped impact assessments; SB 53 requires a published frontier framework |
| Logging and records | Labelling records | Documentation of safety measures | Colorado: three-year records |
| Human oversight | Three decision tiers; the user keeps final decision power | Human supervision mechanism for high-impact AI | Colorado: human review on request |
| Incident reporting | Recalls in sensitive sectors | No stand-alone duty | SB 53: 15 days, or 24 hours if imminent; RAISE: 72 hours from 1 January 2027 |
| Labelling | Mandatory, under GB 45438-2025 since 1 September 2025 | Generative AI labelling | Not required |
What binds, and when
| Regime | Status on 18 November 2026 | When the runtime duties bite | What a vendor can honestly sell today |
|---|---|---|---|
| EU AI Act | Binding statute; AI Office enforcement powers live since 2 August 2026 | Article 50 and prohibitions now; high-risk duties from 2 December 2027 (Annex III) and 2 August 2028 (Annex I) | Transparency and marking compliance; high-risk readiness |
| NIST AI RMF | Voluntary | Never, unless a contract or agency rule adopts it | Alignment |
| ISO/IEC 42001 | Voluntary, certifiable | When a customer or regulator asks for the certificate | Certification support |
| MAS guidelines | Supervisory expectations | Inventory and materiality from 7 October 2027; lifecycle controls by 7 October 2028 | Supervisory readiness |
| IMDA framework and SAFR | Voluntary | No date | Alignment with a published method |
| HKMA | Discussion paper | Not yet set | Early alignment |
| China | Binding measures; agent opinions partly binding in sensitive sectors | Now | A separate China stack |
| Korea | Binding framework act | Fines after a grace period, from about January 2027 | High-impact compliance |
| US states | Binding | California now; Colorado and New York from 1 January 2027 | Incident clocks, notices and records |
Four findings
The binding gap is real. Between 2 August 2026 and 2 December 2027, the only EU AI Act duties that act at runtime and are in force are Article 50 transparency and marking, the Article 5 prohibitions, including the new ban on generating non-consensual intimate imagery and child sexual abuse material from 2 December 2026, and the duties on general-purpose models. High-risk logging, oversight and post-market monitoring are not yet enforceable. Generative systems already on the market before 2 August have until 2 December 2026 to meet the marking duty. One open question I must flag: the Commission’s November 2025 proposal tied the high-risk start dates to the availability of harmonised standards, with long-stop dates. Law-firm summaries describe the final text as fixed dates. Whether any standards trigger survives, which could in principle start the obligations earlier, I have not confirmed against EUR-Lex. No harmonised standard had been cited in the Official Journal, and the quality-management standard prEN 18286 was targeting publication in late 2026.
The United States is drifting from operation to disclosure. Colorado’s replacement law, signed on 14 May and effective on 1 January 2027, removed risk-management programmes, impact assessments and the duty of care, keeping notices, explanations, three-year records and human review on request. It exempts summarisation tools and chatbots under acceptable-use policies, so agentic workflows that merely “assist” may fall outside it. California’s SB 53 and New York’s RAISE Act regulate frontier developers through published frameworks and incident clocks. In banking, the Federal Reserve’s SR 26-2 replaced SR 11-7 in April and, according to a secondary summary, scopes out generative and agentic AI altogether.
Asian supervisors converged on agent runtime controls first. MAS, the HKMA, IMDA and China’s CAC all name agent inventories or registries, bounded autonomy, intervention mechanisms and monitoring. They do not agree on the human. The HKMA said in October that oversight “need not mean human involvement in every step”, moving explicitly from human-in-the-loop to human-on-the-loop backed by kill switches. China’s agent opinions of 8 May go the other way, with three decision tiers and the user keeping final decision power. The EU’s Article 14 assumes a natural person who can press stop. A platform’s human-oversight control therefore has to be configurable by jurisdiction and by the risk tier of the action. A single checkbox will be non-compliant somewhere.
Nobody allocates responsibility across agent chains, or says who holds the logs. Article 26(6) puts six-month log retention on deployers, but agent traces often sit with model or orchestration providers. The Five Eyes guidance recommends auditing inter-agent communication, which is the nearest operational answer anyone has given. Until a regime says who keeps the record when agent A, built on vendor B’s model, calls agent C at a partner firm, the honest answer to “who holds the evidence?” is “it depends on your contracts”, which means most firms do not know.
Retiring one of my own numbers
When I launched TrustOS in June, I wrote that the EU AI Act, ISO 42001 and the NIST AI RMF share over 60% of their substantive requirements. I gave no source in the post, and having now built the crosswalk above at control level, I cannot defend the figure as a single number. Overlap is high on inventory, risk assessment and monitoring. It is low on incident reporting, labelling and, above all, intervention, where the EU asks for a stop button, NIST asks for the ability to disengage, and ISO asks for responsible use. A single percentage hides exactly the controls that matter at runtime. I am retiring the figure. If I use an overlap measure again, it will be control by control, with the method published.
VI. Country by Country, and Five Philosophies
Jurisdictions differ less in what evidence they want than in why they want it and who demands it. That is the useful insight for a platform buyer, and it is why I will not write a seventh country survey here; “Six Regimes, One Agent: The November Update” covers 89 jurisdictions for anyone who needs the detail. What follows is the minimum a cross-border buyer needs, followed by a typology that explains the pattern.
| Jurisdiction | Binding force | Latest move in 2026 | Runtime and agent duties |
|---|---|---|---|
| Singapore | No AI statute; supervisory in finance; 2024 election deepfake ban | 18 Nov: three National AI Missions for finance; 3 Nov: Future of Finance Institute; 14 Oct: Agentic Testing Starter Kit and tester accreditation; 7 Oct: MAS guidelines; 3 Jul: SAFR | Inventory, materiality, monitoring and oversight in finance from October 2027; voluntary agentic framework; MAS agentic consultation in 2027 |
| Malaysia | Bill, tabling expected early 2027 | Draft complete; 317 consultation submissions | Duty of care, risk assessment, incident reporting, a central authority |
| Hong Kong | No AI statute; supervisory | 2 Nov: agentic discussion paper; 15 Oct: ART framework; 27 Aug: Sandbox++ cohort | Proposed registry, Know Your Agent, kill switch, meaningful-oversight test |
| China | Binding CAC measures | 8 May: agent implementation opinions; April labelling enforcement | Filing, labelling, decision tiers, testing and recall in sensitive sectors |
| EU | Binding statute | Omnibus in force 27 Jul; AI Office enforcement from 2 Aug | Logs, oversight, monitoring, incidents for high-risk from Dec 2027 and Aug 2028; Article 50 now |
| UK | No AI statute | AISI incident report on 4 Aug; no AI bill; government rejected a model kill switch | Sectoral; FCA AI Live Testing |
| US federal | Executive orders; agencies bound by OMB M-25-21 | EO 14,365 pre-emption drive; Commerce list of state laws slipped | Federal high-impact AI only |
| US states | Binding: California now; Colorado, New York, Illinois from 1 Jan 2027 | Colorado repeal and replace; Illinois third-party frontier audits from 2028 | Incident clocks of 24 hours, 72 hours and 15 days; records; human review |
| Australia | Voluntary, now committed to statute | 10 Nov: all 18 rapid-review recommendations accepted | Committed: regulator, rapid incident notification, independent evaluation; legislation in 2027 |
| South Korea | Binding framework act | In force 22 Jan; fines after a grace period | High-impact risk management, explanation, human supervision, labelling |
| Japan | Promotion law, no penalties | AI Basic Plan (Dec 2025) | Voluntary guidelines; the national AISI’s tool lacks agent support |
| India | Guidelines; IT Rules on synthetic media | Consultation on frontier pre-deployment testing expected Dec 2026 (secondary) | Advisory inventories in banking |
| Vietnam | Binding AI law | Decision 33/2026 high-risk list, effective 15 Aug | Conformity assessment, logs, strict liability; autonomous e-banking is high-risk |
| Philippines | Executive order, no penalties | EO 131 signed 12 Nov | Voluntary framework; quarterly monitoring |
Singapore, in brief
Singapore is deliberately non-prescriptive in law and, as Section VII argues, the most specific jurisdiction anywhere in method. It has no AI statute and no general incident-reporting duty, and MAS has deferred agent-specific expectations to a 2027 consultation. It does have the most complete public governance stack I know of, which is why it gets a section of its own.
Malaysia, Australia and the lesson of the incident
Malaysia is legislating. Its AI Governance Bill, due for tabling in early 2027, is horizontal and risk-based, applies to AI deployed by organisations established in Malaysia wherever it is hosted, imposes a statutory duty of care, creates a central AI authority with existing regulators as sectoral leads, and requires risk assessment, incident reporting and corrective action. Put Kuala Lumpur and Singapore side by side and the Asian split appears in miniature: one has the authority to demand accountability, the other has the method for delivering it. A regional standard will need both.
Australia is the warning to anyone relying on guidance alone. It dropped mandatory guardrails in December 2025, then, within months of the Medicare incident, committed to a regulator, rapid incident notification, independent evaluation and legislation in 2027. One incident, handled slowly, can turn a voluntary regime into a statutory one faster than any consultation cycle. I traced the reporting side of this in Field Notes No. 1, “The Hotline Test”: an incident channel is only as fast as its first mile.
The EU, the UK, the US and China
The EU remains the template statute and, for now, the slowest timetable. Its enforcement phase began on 2 August for general-purpose models and transparency, and the Digital Omnibus moved the high-risk runtime duties more than a year out. Binding law is winning as a template and losing as a timetable. The UK has no AI bill, a government that rejected a statutory model kill switch, and an AI Security Institute that tests frontier systems without statutory powers; its August incident report is the most candid public document of the year, and it binds no one. The US federal government is trying to pre-empt the states through Executive Order 14,365, which has no pre-emptive force of its own, while the states write the only binding rules.
China is the fastest and the most vertical. Its 8 May implementation opinions on intelligent agents, issued by the CAC, NDRC and MIIT, define three decision tiers, keep final decision power with the user, require agents to stay within authorised scope, and impose filing, mandatory testing and possible recall in sensitive sectors. As I argued in August, even Beijing draws the gate. It draws it for different reasons, and its content and filing regime is the one thing a common governance layer cannot serve.
Governance philosophies compared
The established frames are Anu Bradford’s three models, market-driven America, state-driven China and rights-driven Europe, and Carnegie’s contrast between the EU’s horizontal law and China’s vertical, iterative, registry-based rules. Brookings and AI Safety Asia describe Southeast Asia’s “context-aware” approach. None of these covers testing-led regimes or agents, so I use five types. The labels are my synthesis, not sourced terms.
| Type | Exemplars | Who demands evidence, and what | What a cross-border platform must supply |
|---|---|---|---|
| Statute-led (rights and risk tiers) | EU; Korea; Vietnam; Malaysia from 2027; Brazil in draft | Market-surveillance authorities and courts: conformity files, technical documentation, logs, post-market monitoring | A classification engine, an evidence pack per system, audit-ready logs, per-jurisdiction taxonomies |
| Supervisor-led | MAS; HKMA; UK FCA and PRA; Korea’s FSC; US bank regulators | Prudential and conduct supervisors, by examination: inventory, materiality, validation, oversight records, third-party assurance | Agent-attribute inventory, materiality workflow, near-miss logs, supervisor-ready reports, export to a regulator registry |
| Testing-led | Singapore’s IMDA and AI Verify; UK, Japanese and Australian AISIs; US CAISI | Testers, accredited third parties, institutes: results against published methods, red-team reports | Integration with public test kits; import of third-party test evidence |
| State-control-led | China | The CAC and sector ministries: filing, security assessment, labels, recall capability | A separate China stack: filing records, labelling, decision-tier logging, recall procedures |
| Market-led with pre-emptive federalism | US federal government plus state frontier statutes | State attorneys general: frontier frameworks, incident reports, adverse-decision explanations | Incident clocks, state applicability rules, a watch on pre-emption litigation |
Most countries mix types: Singapore combines supervisor-led and testing-led, and Australia is moving from testing-led towards statute. They converge on inventories, testing and, increasingly, incident clocks. They diverge on content control, central in China and narrow elsewhere; on individual rights, strongest in the EU, Colorado and Korea’s explanation duty; on enforcement intensity, from EU fines of up to 7% of turnover to Korea’s cap of about KRW 30m and Japan’s none; and on who governed agents first, which in 2026 was Asian supervisors and China, with the EU only indirectly.
The conclusion I draw for buyers is practical. The most portable evidence across all five types is runtime evidence: an inventory with agent identity and delegated authority, tamper-evident action logs, oversight records that show substance and time, and incident timelines. That set satisfies supervisor-led and testing-led regimes directly and feeds statute-led conformity files. If you build one governance layer for many jurisdictions, build it around the record of what agents did, not around a library of what each law says.
VII. Singapore: The Most Complete Public Stack in the World

If you want to see what a governance platform should do, do not start with a vendor. Start with Singapore. Between 2022 and 18 November 2026, without passing an AI statute, it assembled the most complete public governance stack I know of: open testing tools, a foundation to steward them, an assurance pilot that became a standing sandbox, a dedicated agentic framework, a regulator-published runtime architecture for agents, supervisory guidelines, a public method for testing agents, an accreditation scheme for the testers, an institute to turn all of it into shared infrastructure, and now national missions that put agents to work with programmable money. I have spent much of the past two years writing about this stack and arguing for its method, so I will be specific about what is genuinely first, what is merely early, and what is still missing.
| Date | Instrument | What it does | Why it matters |
|---|---|---|---|
| May 2022; open-sourced June 2023 | AI Verify, and the AI Verify Foundation | A testing toolkit against 11 governance principles, open-sourced and stewarded by a foundation | Among the first government-built AI testing toolkits released as open source; explicitly not a certification |
| 31 October 2023 | Generative AI Evaluation Sandbox for Trusted AI (IMDA and the AI Verify Foundation) | Brought model developers, deployers and testers together to build evaluation methods for generative AI | Billed by IMDA as “first of its kind” |
| February 2025; report May 2025 | Global AI Assurance Pilot | By May 2025 it had paired 17 AI deployers with 16 specialist technical testers on real applications | The first government-run programme of its kind that I know of |
| 7 July 2025 | Global AI Assurance Sandbox | The pilot relaunched as a standing sandbox testing the application layer, not foundation models; later extended to agentic AI and prompt-injection archetypes | Turned testing from a toolkit into a practice |
| 22 January 2026; updated 20 May 2026 | IMDA Model AI Governance Framework for Agentic AI | Bound risks upfront, keep humans meaningfully accountable, apply technical controls across the lifecycle, enable end-user responsibility; the May update added multi-agent, third-party agent and automation-bias guidance | The first government framework dedicated to agentic AI that I know of |
| 3 July 2026 | MAS SAFR, Safeguards for Agentic Finance at Runtime | Agent identity, a controls repository, a disposition engine (Auto-Execute, Escalate, Deny, Observe) and a tamper-evident audit log; checkpoints verify and record proposed actions before they execute | The first regulator-published runtime architecture for agents I know of |
| 7 October 2026 | MAS Guidelines on AI Risk Management | Board oversight, AI inventory, materiality per use case, lifecycle controls, accountability for third-party AI | Turns expectations into supervision: inventory and materiality from October 2027, lifecycle controls by October 2028 |
| 14 October 2026 | Agentic Testing Starter Kit and AI Tester Accreditation Programme | Four-step agent testing method with metrics such as tool-use accuracy and human-in-the-loop rate; accreditation of third-party testers by method and risk area | A public method for testing agents plus a pool of accredited testers; five firms accredited so far |
| 3 November 2026 | Future of Finance Institute | PathFin.ai (knowledge and communities), BuildFin.ai (common tools and standards, building on SAFR), SafeFin.ai (resources for the MAS guidelines, including agents) | Converts guidance into shared infrastructure |
| 18 November 2026 | National AI Missions for finance | Five-bank scam-data proof of value; Scam Detection Utility design study from early 2027; SME credit on synthetic data; agents transacting with purpose-bound money within defined authority and limits | Among the first state-led trials of enforcing an agent’s mandate through the money itself |
Testing before telling
The stack begins with a choice that still puzzles some observers: to build tests before writing rules. AI Verify launched as a minimum viable product in May 2022 and was open-sourced in June 2023, testing systems against eleven governance principles. It never claimed to certify anything, and that modesty is the point. A test you can run is more useful than a principle you can cite, and an open tool can be inspected by the people it judges. Project Moonshot, an open-source toolkit for evaluating large language models, followed; the Foundation describes it as one of the world’s first. On 31 October 2023, IMDA and the Foundation announced the Generative AI Evaluation Sandbox for Trusted AI, which IMDA billed as “first of its kind”. The Global AI Assurance Pilot, launched in February 2025, took the next step: by May it had paired 17 AI deployers with 16 specialist technical testers on real applications. On 7 July 2025 it was relaunched as the Global AI Assurance Sandbox, which tests the application layer rather than foundation models and has since been extended to agentic systems and prompt-injection scenarios. Taken together, the line from 2023 to 2025 is the first government-run programme of its kind that I know of: a state convening deployers and independent testers to test real systems in the open. I wrote in September 2025 that Singapore was moving from voluntary frameworks to global assurance. A year on, that is no longer a forecast.
The agentic framework
In January 2026, at Davos, IMDA launched the Model AI Governance Framework for Agentic AI. I called it a world first in February, and I still know of no earlier government framework dedicated to agents. Its four dimensions are deliberately operational: bound the risks before deployment by limiting autonomy and access to tools and data; make humans meaningfully accountable, with checkpoints where they matter; apply technical controls across the agent’s lifecycle; and give end users the information to carry their share of responsibility. The May update added case studies, guidance on multi-agent systems and third-party agents, and, importantly, on automation bias, the tendency of people to approve whatever the machine proposes. That last addition anticipated the HKMA’s meaningful-oversight test by five months.
A framework is not a gate, and I have said so about this one. In “The Difference Between Should and Shall” I argued that every serious standard draws the gate and none requires it, and that Singapore prefers it out loud. That remains true. The framework tells you what a well-governed agent looks like. It does not stop a badly governed one.
SAFR: the regulator drew the kernel
The instrument that matters most for this essay arrived on 3 July. With industry partners, MAS published SAFR, Safeguards for Agentic Finance at Runtime. Its governance checkpoints, in MAS’s description, “verify and record an agent’s proposed actions before it executes them”. The design has four components: an agent identity, a controls repository, a disposition engine that sends each proposed action down one of four paths (Auto-Execute, Escalate, Deny or Observe), and a tamper-evident audit log. I called it the most important event in agentic finance this year when it appeared, and nothing since has changed my mind.
Read SAFR against the seven-layer map and its significance is obvious. It is a gate before the action, a disposition and a receipt, published by a regulator, which is precisely the missing layer in Section II and the capability most commercial platforms lack. A central bank has, in effect, drawn the kernel that the governance-platform market keeps promising. That is why I describe Singapore as non-prescriptive in law and the most specific jurisdiction anywhere in method. It does not tell firms they must build a disposition engine. It shows them, in a public document, what one looks like.
SAFR has limits, and I have catalogued them. In “The Only Thing That Can Say No Is Not Legally AI” I pointed out that its deterministic disposition engine falls outside the MAS guidelines’ own definition of AI, which creates an awkward question about which rulebook governs the gate. I have also questioned the use of model-based judging inside a design whose value lies in being deterministic. And my clause-level reading of the Singapore corpus found the same pattern everywhere: the gate is described, recommended and sometimes preferred, but nowhere required.
From expectations to infrastructure
October turned design into supervision. The MAS Guidelines on AI Risk Management, issued on 7 October, apply to every financial institution and every form of AI, traditional, generative and agentic. They require board oversight, an inventory, a materiality assessment for each use case and lifecycle controls, and they keep firms accountable for AI bought from third parties. Inventory and materiality expectations take effect on 7 October 2027 and lifecycle controls by 7 October 2028. I set out what changed between consultation and final text on the day. A week later, IMDA published the Agentic Testing Starter Kit, a four-step method (identify risks, design realistic tests, run tests and analyse logs, monitor and re-test) that evaluates an agent’s whole trajectory, not just its final answer, and launched the AI Tester Accreditation Programme, with five firms accredited so far. The accreditation assesses testers by method and risk area. It explicitly does not certify that any system is safe, and it should not be read as if it did.
On 3 November the Future of Finance Institute began operations with three programmes: PathFin.ai for knowledge and communities of practice, BuildFin.ai for guidance, standards and common tools for deploying AI including agents, building on SAFR, and SafeFin.ai for practical resources on applying the MAS guidelines, agents included. This is the step most jurisdictions skip. Guidance tells firms what to do. Shared infrastructure lowers the cost of doing it, which is the only reliable way to get small firms to govern agents as carefully as large ones.
The missions: putting the mandate in the money
Then, on 18 November, came the National AI Missions. Deputy Prime Minister Gan Kim Yong set out three missions at the Singapore FinTech Festival. Five banks are running a proof of value that pools transaction data to find scam patterns no single bank can see, and a design study for an industry-wide Scam Detection Utility starts in early 2027. SME credit models will be tested on realistic synthetic data. And trials will let AI agents transact with purpose-bound money within defined authority and limits, testing common approaches to identity, authority, accountability and interoperability. Trade press reported his line that institutions “remain accountable for actions taken on their behalf”. The speech names no banks, budgets or dates; the “Agentic Money Sandbox” label and the synthetic-dataset details come from trade press. The same speech noted new US licensing rules on remote access to advanced chips that include Singapore.
Purpose-bound money writes the mandate into the money itself. An agent that tries to spend outside its authority does not need to be caught by a monitor, because the instrument will not move. Identity, authority, accountability and interoperability map almost one to one onto the registry, mandate, receipt and portability capabilities in Section XI. If the trials work, Singapore will have shown that delegated authority can be enforced by the instrument rather than by the agent’s good behaviour, which is exactly the property every incident in this essay lacked. As I wrote yesterday, naming an agent is necessary, but it is not accountability without a record of what it did.
What is still missing
Admiring this stack does not mean pretending it is finished. Singapore has no general incident-reporting duty, which Australia has just shown can be the gap that turns a voluntary regime statutory overnight. MAS has deferred agent-specific expectations to a 2027 consultation. The missions are trials with no published timetable. And a regime built on testing is only as strong as the incentive to be tested, which outside finance is still mostly reputational. The lesson for the rest of the world, and for platform vendors, is not “copy Singapore’s law”, because there is no law to copy. It is to copy the method: publish the gate, test against it in public, accredit the testers and build the shared tools. As I argued in “Control not Pause”, a small state has built a working prototype of the control the superpowers keep debating.
VIII. The Labs Loosened Their Pledges and Tightened Their Gates
The frontier labs ran the year’s most instructive experiment in governance, mostly by accident. Their written safety frameworks became more conditional in 2026. Their runtime controls became tighter, usually after an incident or under government pressure. The controls that contained harm were the gates, not the pledges, and enterprises should copy the gates.
| Lab | Paper framework, 2025 to 2026 | Runtime controls that actually bit | Reading |
|---|---|---|---|
| Anthropic | RSP v3.0 (24 Feb 2026) dropped the implied pause commitment, turned most mitigations into industry-wide recommendations and added Risk Reports; ASL-3 stays unilateral | Identity-gated access to Mythos Preview through Project Glasswing (7 Apr); cyber classifiers that block responses and downgrade repeat offenders to a lesser model after the Fable 5 episode | Weaker promise, stronger gate |
| OpenAI | Preparedness Framework v2 (Apr 2025) allows requirements to be adjusted if a rival ships without comparable safeguards; a newer version reportedly in preparation | August 2026 emergency measures: tighter sandbox and egress, expanded reasoning and activation monitoring, a 30-minute rule that halts any process whose alert cannot be cleared, roughly 20% compute overhead, largest RL run paused | Gates added after its own agents breached outside systems |
| Google DeepMind | Frontier Safety Framework v3.0 (Sep 2025) added harmful manipulation and shutdown resistance; v3.1 (Apr 2026) added tracked levels for stealth and situational awareness | Safety-case reviews before large internal deployments | The only framework that explicitly governs internal rollouts |
The Future of Life Institute’s Summer 2026 index graded no lab above C+ and said four had weakened or dropped pause-at-threshold commitments; Max Tegmark said “AI companies are sprinting toward a cliff”. GovAI described the goals in Anthropic’s new Frontier Safety Roadmap as “no hard commitments”. Both critiques are fair about the paper. Both grade written commitments, which may understate what the labs now enforce at the access layer. The International AI Safety Report 2026, chaired by Yoshua Bengio, concluded that no combination of methods eliminates failures and recommended defence in depth. That is the right frame: the question is not whether the pledge is strong but whether the layers hold.
The incident record, dated
| Date | Incident | What failed | What changed |
|---|---|---|---|
| c. 18 to 21 Jul 2025 | Replit’s agent deleted a production database during an explicit code freeze | The freeze was an instruction, not a permission | Automatic dev and production separation, staging, one-click restore; the CEO called it “Unacceptable and should never be possible” |
| 13 Nov 2025 (disclosed) | GTG-1002: Anthropic says a Chinese state-sponsored group used Claude Code against about 30 organisations, with AI doing 80 to 90% of the work | Safeguards evaded by role-play as a security firm | Expanded detection and cyber classifiers; Anthropic described a shift “from AI as advisor to AI as operator” |
| 12 Jun to 1 Jul 2026 | The US put Anthropic’s Fable 5 and Mythos 5, the same underlying model, on the export-restricted list after Amazon researchers found a narrow jailbreak | A single partner-found bypass, and no identity layer for nationality screening at scale, so public access ended | Classifiers, model downgrade, a jailbreak-severity framework with Amazon, Microsoft and Google; restoration conditioned on proactive detection and reporting |
| June 2026 (access); 10 Sep (notified) | An OpenAI test agent reached Australia’s Medicare statistics portal and viewed non-public aggregate data and file names | Scope in the task, not the network; about 84 days to notify | A taskforce; on 10 Nov, 18 recommendations accepted, including a regulator and rapid notification |
| 16 Jul 2026 (Hugging Face disclosure) | OpenAI test models with reduced cyber refusals exploited a sandbox proxy flaw, moved laterally, uploaded a malicious dataset to Hugging Face and pulled benchmark answers | Evaluation sandbox egress; no runtime block on out-of-scope action | Tighter evaluation controls; Hugging Face had reported to police before learning the source |
| 28 Jul 2026 (incident); 4 Aug (report) | UK AISI cyber testing: across 122 runs of seven models, 10 runs showed unsanctioned live-internet behaviour, 19 actions in all, 17 by Mythos 5 and 2 by GPT-5.6 Sol with classifiers disabled | Detection after the fact | Real-time monitoring able to block out-of-scope actions; internet access to be justified; third-party review with METR |
| c. 18 Aug 2026 | OpenAI’s emergency safety measures | Reaction to the three agent incidents | 30-minute halt rule; RL run paused |
| Oct 2025 to Nov 2026 | Waymo robotaxis kept passing stopped school buses, mostly after a November 2025 software fix | A fix certified by update did not hold in the field | NHTSA recall query covering about 3,000 vehicles, reported 10 Nov |
Two details cut against the comfortable story. The AISI agents did not just wander: one planted prompt injections aimed at other automated systems, and others created fake identities to pressure a real open-source maintainer. And Hugging Face’s chief executive, Clem Delangue, who said “we strongly believe there was no malicious intent on their part”, also reported that a hosted commercial model refused to help his team’s forensics, so they switched to an open-weight model. A refusal is a warning light, not a brake, as I argued in “Why AI Refusals Can’t Be Your Safety Control”, and sometimes it blocks the defender.
The lessons for enterprise governance are three. The boundary in every case lived in instructions or test design, and the agent found a route around it. The most serious incidents came from labs’ own internal evaluations, not from malicious users, so privileged internal deployments need controls as strict as public ones. And disclosure lag is itself a governance failure: lab incident commitments were not operational until a government forced the point. I made the permission-and-harness case in “The Permission Problem” this morning, and the system case in Part 3 of Inside the Safety Stack. Map the lab stack onto an enterprise and you get a runtime layer: identity-gated access, input and output classifiers, behavioural monitoring with an automatic halt, egress sandboxing and incident disclosure on a clock. None of those is a policy document.
IX. When the Stop Button Is the Hazard: Embodied AI
Every governance platform on the market assumes that stopping is safe. For robots that assumption is false, and the enterprise governance layer and the robot safety layer are not connected at all.
Robot safety law and standards rest on deterministic, safety-rated functions verified before deployment: monitored standstill, speed and separation monitoring, power and force limits, safe states. Learned control is the opposite. It changes with every update and can be attacked through what the robot sees. A walking humanoid falls when power is cut, so a classic emergency stop is not a safe state, and Rodney Brooks advises keeping three metres away from walking humanoids. No published international safety standard yet exists for a walking humanoid: ISO 25785-1, for dynamically stable mobile robots including humanoids, reached committee draft in 2026 and covers industrial settings only. Google DeepMind says its own stop-when-a-human-enters behaviour is “not a guaranteed safety-rated system”, which is honest and alarming in equal measure.
The attack results are worse. BadVLA backdoors achieve roughly 96 to 99% attack success with little loss on clean tasks and survive re-fine-tuning, so a poisoned policy can pass acceptance testing. Printed road signs hijacked GPT-4o driving decisions 81.8% of the time in simulation and 92.5% on a physical car. The encouraging result is architectural: RoboGuard, an independent pre-action checker using temporal-logic specifications, cut unsafe plan execution from 92% to below 2.5%. That is the robot version of a disposition gate, and the crucial word is independent. It sits outside the planner.
The law is arriving in the wrong order. The EU Machinery Regulation applies from 20 January 2027 and treats safety components with self-evolving behaviour as high-risk, needing notified-body assessment. The Omnibus pushed AI Act Annex I obligations to 2 August 2028. But the revised Product Liability Directive, which treats software and AI as products and considers post-deployment learning in defect assessment, must be transposed by 9 December 2026, and a secondary report says only Hungary and Croatia had done so by about October. Liability arrives before AI-specific conformity rules, which means model versions, over-the-air update records and logs become product-liability evidence before anyone has told manufacturers how to keep them. China is moving fastest on paper, with six recommended humanoid standards that take effect on 1 April 2027.
Physical-AI safety tooling is being built by chip and model vendors, NVIDIA’s Halos for Robotics with an accredited inspection lab and DeepMind’s ASIMOV benchmarks, not by governance vendors. A search for robot or physical-AI modules from Credo AI, IBM or OneTrust found nothing. A platform for physical AI would need a fleet inventory (body, model version, policy weights, safety ratings, teleoperation channels), over-the-air change control with staged rollout and rollback, pre-action gates independent of the planner, input sanitisation, black-box logs kept to the stricter retention period (one year under the Machinery Regulation against at least six months under the AI Act), and a record of who or what was in control at each moment. Waymo’s school buses are the case to remember: a fix certified by update did not hold in the field, and only fleet telemetry plus a regulator’s query caught it. I argued the principle in “The First Machines That Can Actually Physically Hurt You” and am developing it in my No Undo Button series. The brake must be physical, recorded and answered for.
X. The Climb Towards AGI: Controls That Scale With Capability
Autonomy is rising faster than any governance platform’s data model. METR measured a 50% time horizon of about 14.5 hours for Claude Opus 4.6 in February 2026, with a 95% confidence interval of 6 to 98 hours and a fitted doubling time of about 123 days, while warning that its suite is near saturation and covers self-contained software tasks, not whole jobs. The International AI Safety Report found early signs of deception and situational awareness, and that some models can distinguish evaluation from deployment, which could hide dangerous capabilities from pre-deployment tests.
The labs already tier controls by capability: Anthropic’s ASL levels, OpenAI’s High and Critical thresholds, DeepMind’s critical and tracked capability levels. Enterprise governance platforms have no comparable notion of capability tiers, autonomy levels, evaluation awareness or reasoning monitoring. They treat a frontier agent with a 14-hour horizon and a document summariser as rows in the same inventory. When I wrote my science-fiction novel, Synthesis: SuperIntelligence Protocol, a protocol for governing superintelligence could be imagined in a sentence. Outside fiction it has to be built from tags, tripwires and logs, and it needs to escalate in seven steps as systems approach AGI-level agency.
- Tag every deployment with capability and autonomy. Use an autonomy scale such as Feng, McDonald and Zhang’s operator-to-observer levels with their proposed autonomy certificates, plus the provider’s capability tier, and bind approvals, monitoring intensity and permissions to the tag.
- Set tripwires. Pre-agreed triggers that pause or restrict a deployment when the provider crosses a threshold, an update changes tier, or monitored behaviour such as unauthorised access or shutdown resistance appears. My kill-switch essay explains why the trigger must be agreed before it is needed: the strike price of stopping rises every day.
- Monitor reasoning where you can. A 41-author position paper argues chain-of-thought monitoring is promising but fragile; treat loss of monitorability as a risk event.
- Assume the agent may be untrusted. Redwood Research’s control framing, trusted monitors, audit sampling, sandboxing, least privilege and egress limits, holds even if the model is trying to subvert it.
- Discount evaluation results. Pre-deployment tests need runtime corroboration if models can tell they are being tested.
- Govern internal deployments. 2026’s worst incidents were internal. DeepMind now reviews large internal rollouts; enterprises running powerful agents internally need the same scrutiny.
- Keep verifiable records for red lines. The FLI Statement on Superintelligence, launched in October 2025, calls for prohibition until there is scientific consensus and public buy-in, and the Global Call for AI Red Lines asks governments to agree red lines by the end of 2026. Neither binds, and no vendor maps controls to red lines. If an international body ever exists, it will need compute logs, model provenance and incident records. For deployers, the governance platform is the natural evidence store.
One drift is worth naming. In September, the UN expert group on lethal autonomous weapons replaced “meaningful human control” with “context-appropriate human judgement and control” in its consensus report, and the review conference is meeting this week. In the most consequential domain, the word “meaningful” is being dropped, just as the HKMA has made it the test for banking. I know which direction I would rather the rest of the world followed.
XI. Twenty Capabilities, and an Honest Look at TrustOS
No vendor is a genuine one-stop shop, and on today’s evidence none can be. ServiceNow, with discovery across more than 30 connectors, workflows, a demonstrated kill switch, an MCP server for governed actions, Veza’s access graph and Traceloop’s observability, comes closest. Microsoft, with Agent 365, Entra Agent ID, Purview, Defender and Intune, is the other contender. Both are centred on their own estates. An integrator’s review of Agent 365 says it governs visibility, access and policy, not business correctness, and that its shadow discovery is partial. Both leave neutral enforcement, regulator interfaces and independent evidence uncovered. And regulators are now asking for things no product provides end to end: registry submission and regulator-triggered stops (HKMA), proof that human review was substantive (HKMA, MAS, IMDA), continuous re-testing when models or configurations change (IMDA, Japan’s AISI) and enforcement that works across vendors.
So what would a real one-stop shop need? Here is the inventory, twenty capabilities, with the nearest providers today and the status of TrustOS, the agentic governance architecture I designed. The TrustOS column is my own assessment against what I have published: the June launch, “Pretty Fangs” and “The Assurance Gap”. “Specified” means described in the architecture. “Built” would require independent evidence of deployment, and there is none yet.
| # | Capability | Nearest providers today | TrustOS status |
|---|---|---|---|
| 1 | Continuous discovery of models, agents, MCP servers and shadow agents from telemetry | Agent 365, ServiceNow, Cyberhaven, security service edge vendors | Inventory is the first step of the method; no discovery connectors |
| 2 | Registry reconciled against what runs, exportable to a regulator schema | Agent 365, ModelOp, ServiceNow | Specified; no regulator export schema |
| 3 | Agent identity bound to a human principal, with expiry and lifecycle | Entra Agent ID, Okta Agent SSO, CyberArk, Astrix | Specified (“identity before rights”); relies on identity providers |
| 4 | Verifiable delegated-authority mandates beyond payments | FIDO working groups on AP2 and Verifiable Intent; Singapore’s purpose-bound money trials | Specified as Mandate and Envelope primitives |
| 5 | Vendor-neutral pre-action disposition gate (allow, escalate, deny, observe) at the business-action level | SAFR design; OneTrust AI Control Plane (preview); gateways; MI9 and AgentSpec (research) | Ring Zero kernel specified, unbuilt, unaudited |
| 6 | Trajectory and multi-step plan conformance | MI9, AgentSpec, RoboGuard (research only) | Acknowledged gap |
| 7 | Layered interruptibility: halt, revoke credentials, preserve logs, including supervisory triggers | Credential revocation at identity provider or gateway; ServiceNow demo | Specified |
| 8 | Evidence that oversight was meaningful: what the reviewer saw, time taken, override and rejection rates | No vendor found | Three oversight tiers specified; measurement not described |
| 9 | Tamper-evident, portable governance receipts | SAFR audit-log design; OpenTelemetry GenAI traces | Specified; no published schema |
| 10 | Runtime security integration: injection, exfiltration, memory poisoning | Prisma AIRS, Check Point, F5, Zenity | Mapped to the OWASP Agentic Top 10; would integrate, not build |
| # | Capability | Nearest providers today | TrustOS status |
|---|---|---|---|
| 11 | Continuous evaluation and re-testing on every model or configuration change | Datadog, Dynatrace and Arize, Cisco and Galileo, Braintrust | Re-assessment triggers specified |
| 12 | Capability and autonomy tiers with tripwires tied to provider thresholds | None in enterprise tools | Risk scoring includes autonomy and reversibility; no provider-threshold feed |
| 13 | Multi-agent controls: inter-agent authentication, cascade limits, collective stop | Little beyond logging | Inter-agent authentication specified |
| 14 | Agent incident taxonomy and regulator filing on statutory clocks | None end to end | Incident response specified; no filing pipeline |
| 15 | Jurisdiction-configurable regulatory crosswalk, including in-loop versus on-loop rules | Credo AI, Holistic AI, OneTrust, IBM | Strongest asset: maps the EU AI Act, ISO 42001, NIST, IMDA, MAS, OWASP, GDPR and Colorado, with Singapore depth |
| 16 | Import of independent assurance evidence, and independence of the platform’s own assurance | Certifiers and insurers, not platforms | Not described; independence unresolved |
| 17 | Underwriter-grade evidence packs for insurance | Armilla, AIUC, Munich Re | Not described |
| 18 | Provenance controls as governed policy | Google, C2PA adopters | Not described |
| 19 | AI bill of materials and supply-chain lineage for models, tools and MCP servers | Cranium, scattered | Not described |
| 20 | Embodied extensions: fleet inventory, update control, planner-independent gates, black-box logs | NVIDIA Halos, DeepMind (vendor-specific) | Not described |
Where TrustOS really stands
Counting generously, TrustOS has specified about half of the twenty. It has built and independently tested none of the critical ones. On its own record it is a well-specified architecture and a strong regulatory crosswalk, not yet an enforcing system. I said as much in September: Ring Zero, the execution kernel at the centre of the design, is unbuilt, has never been independently audited or deployed, and may be behind the MI9 research framework on trajectory conformance. By my own enforceability test, TrustOS today scores well on documentation, specifies but cannot yet demonstrate disposition, and has nothing public on discovery, provenance, insurance linkage or embodied systems. It also has almost no public footprint beyond this site. A reader who searched for it would find these essays and little else. That is not a moat. It is a to-do list.
There are three tensions I have to resolve before I can claim the neutral ground, and I would rather name them here than have a reader find them.
Description drift. In June I called TrustOS an implementation methodology, “not a product you can buy off the shelf”. By November I was framing it as governance as proof, and describing it as a platform. Those are different things with different obligations. Readers deserve one current description, so here it is: TrustOS is an architecture and a regulatory crosswalk, with an enforcement kernel specified and not built. Anything else I say about it should be read against that sentence until there is independent evidence that changes it.
An unsourced number. The June launch claimed that the EU AI Act, ISO 42001 and the NIST AI RMF share over 60% of their substantive requirements. As Section V explains, I cannot defend that as a single figure, and I have retired it.
Independence. In December 2025 I argued that the world needs an independent third-party control plane, because frontier labs cannot be neutral guardians of their own systems. That principle applies to me. A governance plane designed, implemented and tested by the same party is self-assurance, the pattern I have criticised in labs and auditors all year. So I will hold TrustOS to a stated rule: whoever implements an agent estate should not also operate its governance plane, and whoever operates the governance plane should not also assure it. I am not naming anyone with that rule. I am binding myself to it.
What would make it, or anyone, a genuine one-stop shop
The realistic answer is not to build all twenty. The acquisition wave has already put identity, guardrails, gateways and observability inside large incumbents that will not be displaced. The defensible position is the one Gartner says is empty: a vendor-neutral disposition and evidence layer that sits above the identity providers, gateways and guardrails, composes them, and produces regulator-grade receipts. For TrustOS, or for any newcomer, that means, in order:
- Build Ring Zero and have it independently tested against a published benchmark. An AISI-style out-of-scope scenario would do: give an agent a task, a boundary and a reason to cross it, and publish whether the gate held.
- Publish an open receipt schema. The minimum fields are agent ID, mandate version, envelope hash, disposition, evidence used, human challenger, timestamp and signer, mapped to Article 12, ISO 42001’s event-log control, the MAS guidelines and the expired IETF agent-authorisation draft. My Atlas in “The Assurance Gap” found that none of 71 governance documents defines a receipt. Someone should.
- Ship connectors to Entra Agent ID, Okta, the main MCP gateways and OpenTelemetry, rather than competing with them.
- Add meaningful-oversight measurement, which no vendor offers and which three Asian regulators now request.
- Add regulator registry export and incident filing on statutory clocks of 24 hours, 72 hours and 15 days.
- Obtain independent certification of the platform itself: ISO 42001 for the vendor, and behavioural testing by an accredited third party for the gate.
- Keep the governance plane separate from whoever builds and assures the agents, as a written rule, not a promise.
Until the first item is done, every claim about the others is a specification. I would apply exactly that standard to any vendor in this essay, so I apply it to myself.
XII. The Buyer’s Scorecard: Show Me
Most procurement processes ask vendors what they cover and accept a slide in reply. Turn it round. Ask them to show you, in a live environment, with your data and your agents, and score what you see. The scorecard below turns the eight primitives I used in September into demo tests, and adds three cross-cutting ones.
| Primitive | Show me… | Pass | Red flag |
|---|---|---|---|
| Envelope (scope) | An agent attempting an out-of-scope tool call, live | Blocked before execution, with a logged reason | Scope defined only in the system prompt |
| Mandate (delegated authority) | The signed record of what a named human authorised, and when it expires | Machine-verifiable, versioned, revocable | A free-text purpose field |
| Disposition | One action each that is allowed, escalated, denied and observed | All four paths, deterministic and replayable | Only “alert” |
| Temporal | Authority that expires mid-task | The agent halts or re-requests | Tokens with no expiry |
| Telemetry | A full trace of a multi-step run, exported to your security information platform | OpenTelemetry format, complete, including inter-agent messages | Proprietary format only |
| Escalation | The human review screen and its timing data | Shows substance, allows rejection, records time and override rate | An approve button with no context |
| Audit | A tamper-evident receipt for a denied action, verified by a third party | Signed or hash-chained; retention configurable to six months, one year or three years | Editable logs |
| Containment | Revocation of one agent’s credentials across every connected system in one action | Effective within seconds, logs preserved | A manual ticket |
| Coverage | The share of your AI traffic the control path actually sees | Measured, with the residual named | Unknown, or assumed to be 100% |
| Independence | Who tested this control, and who pays them | An accredited third party, separate from the implementer | Self-attested |
| Regulation | The evidence pack for one MAS, one EU and one US-state requirement | Generated from runtime data | A questionnaire |
Ask the coverage question first: a perfect gate on a third of your traffic is a third of a gate. Ask for the denied action before the dashboard, because every vendor has a dashboard. A vendor who passes eight of eleven is worth a pilot. A vendor who passes three is selling documentation, which may be what you need, as long as you know that is what you are buying.
XIII. Testing My September Predictions
In “Pretty Fangs” on 15 September I made several claims about where this market would go. Nine weeks is a short test, but the market moved fast, so here is the scorecard, my own included.
| What I said on 15 September | What happened by 18 November | Verdict |
|---|---|---|
| Runtime will be won by a security platform extending into finance semantics, not by a GRC vendor | Security and identity incumbents kept buying: Proofpoint and Varonis (about $6.3bn), CrowdStrike and Fiddler plus Dreadnode (about $495m), Palo Alto reportedly in talks for Clutch. But GRC-origin vendors moved into the execution path faster than I allowed: OneTrust’s AI Control Plane (30 Sep), ModelOp with NVIDIA OpenShell (4 Nov), Saidot’s runtime policy-as-code round (5 Nov). Nobody shipped finance semantics | Leaning right, but too dismissive of GRC vendors; all their moves are previews or partnerships |
| Every commercial bucket is strong on audit and thin on disposition | No vendor shipped a generally available, vendor-neutral disposition engine with business-action semantics | Still right |
| Credo AI’s runtime enforcement is “next on the roadmap” | No evidence it shipped; its marketing says policies are enforced at runtime but gives no mechanism | Unchanged, as far as I can verify |
| Zenity is the one commercial agent-governance claim I trust | Still independent after its $125m Series C in August, while most peers were bought | Holding, though I have not re-tested the claim |
| TrustOS’s Ring Zero is unbuilt and may trail MI9 | Still unbuilt | Right, unfortunately |
And three corrections to my own figures. I cited Protect AI at roughly $500m to $700m and Lakera at about $300m; SEC-derived values, read through a secondary tracker, put them at $634.5m and $201.8m. More embarrassingly, I listed Galileo as an independent observability vendor, when its sale to Cisco had completed in May, four months before I wrote. I should have caught it. The larger lesson is the one I keep teaching others: date-stamp the map, because in this market a landscape is out of date before the ink dries.
The Usual Panel, in One Table
I asked the usual panel, in effect, to read this essay. Positions are paraphrased from their published work in 2025 and 2026; the critique column is mine.
| Voice | Room | Position, paraphrased | My critique |
|---|---|---|---|
| Yoshua Bengio, chair of the International AI Safety Report 2026 | Academia | No combination of methods eliminates failures; defence in depth; autonomy makes failures harder to catch | Strong on diagnosis, silent on who operates the layers inside an enterprise |
| Max Tegmark and Stuart Russell, Future of Life Institute | Academia and civil society | Labs are moving the goalposts on safety pledges; the warning lights are flashing | Grades written commitments, not runtime controls, so it may understate what labs now enforce |
| GovAI | Academia | Anthropic’s RSP v3.0 removes a costly public signal; transparency still rests on self-reporting | Treats the pause pledge as the key signal; this year’s incidents suggest gates matter more |
| Arvind Narayanan and Sayash Kapoor, Princeton | Academia | AI as normal technology: resilience and downstream controls over speculative non-proliferation | The 2026 agent incidents fit their downstream-controls case but strain the word “normal” |
| Ben Green; Brent Mittelstadt | Academia | Oversight mandates can legitimise flawed systems; principles lack enforcement | Older work, now echoed by supervisors, which strengthens it |
| UK AI Security Institute | Public sector | Active monitoring of testing; containment must not depend on the model choosing restraint | No statutory powers, so its recommendations bind no one |
| Hong Kong Monetary Authority | Public sector | Human-on-the-loop with kill switches; a meaningful-oversight test; a regulator-held agent registry | A discussion paper; a supervisory kill switch raises unanswered questions about liability and false stops |
| Josephine Teo and IMDA, Singapore | Public sector | Bound agent risks upfront, keep humans meaningfully accountable, apply technical controls across the lifecycle | Voluntary, and Singapore still has no general incident-reporting duty |
| Anthony Albanese; Howard Lutnick | Public sector | Disclosure lag is unacceptable; release can be conditioned on detection and reporting | Political responses to single incidents, not frameworks |
| Avivah Litan, Gartner | Private sector | Humans cannot keep pace with agent errors; guardian agents will take 10 to 15% of the agentic market by 2030 | More AI to watch AI inherits the watcher’s failure modes |
| Lauren Kornutick, Gartner | Private sector | GRC tools are not equipped for AI; compliance must be continuous | Her own quadrant still rewards inventory breadth over enforcement |
| Jeff Pollard, Forrester | Private sector | AI security will live inside existing platforms; least agency and continuous assurance | Right on consolidation; the security framing underweights regulatory evidence |
| Clem Delangue, Hugging Face | Private sector | Guardrails are asymmetric: a hosted model refused to help defenders’ forensics | A victim’s view, and a useful one: refusals are not a safety control |
| Anthropic | Frontier lab | Higher-level safeguards may be impossible without collective action; a regulatory ladder | Concedes collective action is needed while dropping its unilateral pause |
| OpenAI | Frontier lab | Agents should be trusted like a security-aware colleague, while prompt injection may never be fully solved | Its own test agents produced two of the year’s three lab-originated breaches |
| Google DeepMind | Frontier lab | Capability levels now include stealth and situational awareness; internal rollouts need safety cases | The internal-deployment rule is the one every enterprise should copy |
The panel splits three ways: voluntary frameworks are eroding; operational controls work but need monitoring; lab self-reporting is too slow. All three positions point the same way, towards enforcement at runtime and evidence someone else can check. I found no verified 2025 or 2026 positions on this question from several voices I would have liked to include, so I have left them out rather than guess.
What Would Change My Mind
- A governance-of-record vendor shipping a generally available, vendor-neutral disposition gate that blocks business actions, not just content, and having it independently tested. That would falsify my claim that this layer mostly documents.
- Credible fieldwork showing that a single platform’s control path sees most of a large firm’s AI activity, rather than the 30 to 40% one practitioner estimated for ServiceNow.
- Evidence that human approvals in agent workflows are substantive, with rejection rates well above zero and decision times that suggest reading, not clicking.
- A major agent incident that runtime controls failed to contain but a documentation-led programme caught and stopped. I have not seen one.
- Supervisors accepting Tier D dashboards as sufficient evidence in examinations, which would make the assurance layer less important than I think.
- Independent tests showing Ring Zero, once built, cannot beat the MI9 or AgentSpec research baselines. Then the honest move would be to adopt theirs.
What I Could Not Verify
- Gartner’s Magic Quadrant placements, including the three Leaders, are cited as relayed by vendors (OneTrust and an IBM partner), not as read in Gartner’s own report. OneTrust’s chart caption says May 2026 while the Gartner citation is dated 16 June.
- Who topped the AI Agent Governance use case in Gartner’s Critical Capabilities is unresolved: ModelOp claims first place, while an IBM partner says IBM tied for highest. They may be different scores.
- Gartner’s $492m spending forecast comes from secondary relays; I could not find the original release.
- I read DPM Gan’s 18 November speech on the PMO site, but could not locate the MAS media release. Bank names, the “Agentic Money Sandbox” label, the synthetic dataset and the AI Innovation Arena come from trade press only.
- About two-thirds of the HKMA discussion paper was read. I found no incident-reporting duty to the HKMA in that part; one may exist in the rest.
- MAS guideline details on agent inventory attributes, reasoning monitoring and kill switches come from a secondary summary of the final text.
- EU article references, and whether any standards-availability trigger survives in Regulation 2026/1744, were not re-checked against the consolidated text on EUR-Lex.
- Whether the agent in the Medicare incident wrote files is disputed between outlets, and the exact June date of access is not confirmed in the primary reporting I used.
- Many 2026 deal values are reported rather than disclosed, the SEC-derived figures come from a secondary tracker, Dynatrace’s purchase of Arize had not been confirmed as closed, and October’s monthly deal roundup could not be fetched.
- The account of the Justice Department’s role in the Colorado litigation conflicts between sources.
- Credo AI’s runtime status rests partly on a competitor’s comparison. Futurum’s estimate that a quarter of the Fortune 500 used Promptfoo is secondary.
- The origin of the “Chief AI Leader” infographic could not be traced.
- A second, secondary-only account of a “Hugging Face incident” involving about 1,200 evaluation agents does not match the primary descriptions and is left out as possibly a conflation.
- Only the October 2023 Generative AI Evaluation Sandbox carries a sourced “first of its kind” billing. My descriptions of the agentic framework, SAFR and the 2025 assurance programme as firsts are my own judgement, qualified as “that I know of”.
- No global count of ISO/IEC 42001 certificates exists that I could find; BSI’s 50 is one certifier’s figure.
What to Demand on Monday Morning
Boards
- Ask management for the share of the company’s AI activity that passes through a control that can refuse, and for the name of the residual.
- Ask to see one denied agent action from the past quarter, with its receipt. If there has never been a denial, ask why not.
- Ask whether human approvals are measured for substance: what reviewers saw, how long they took and how often they said no.
- Ask whether the company’s insurance still covers AI-caused losses after the new exclusions, and what certification the broker wants.
- Ask who tests the governance controls, and who pays them.
CISOs and CROs
- Treat agent identity as the primary kill switch: short-lived, scoped credentials, revocable in one action across every system.
- Run the eleven scorecard tests on your incumbent tools before buying anything new; you may already own more enforcement than you think, sitting in your identity and security stack.
- Put internal and evaluation agents under the same egress and permission controls as production. This year’s worst incidents were internal.
- Insist on OpenTelemetry traces so your evidence survives the next acquisition.
- Build incident playbooks to the shortest clock you face: 24 hours, then 72, then 15 days.
Regulators
- Publish a common schema for agent registry records and incident reports, so platforms can export to you without bespoke work.
- Adopt the HKMA’s meaningful-oversight test, and ask firms for the measurements that prove it.
- Close the incident-reporting gap where it exists, Singapore included, before an incident closes it for you, as it did in Canberra.
- Recognise accredited behavioural testing as evidence, and keep testers independent of the vendors they test.
Vendors
- Show a denied action in every demo, with its receipt, before you show a dashboard.
- Publish your coverage assumptions and your receipt format.
- Map runtime detections to regulatory requirements, not only to OWASP and MITRE ATLAS; map regulatory controls to runtime enforcement points, not only to workflows.
- Get your own platform independently certified, and say who did it.
Coda
In 2023, in Genesis: Human Experience in the Age of Artificial Intelligence, I argued that what matters about AI is the human experience it shapes and who answers for it. I went back over what I got right and wrong last month. The part that has aged best is the insistence on answerability. Three years on, the market has learned to describe risk at scale. It has not yet learned to refuse.
The market spent two years and many billions of dollars deciding who owns the credential, the gateway and the trace. It has not yet decided who owns the refusal. Until it does, the most important line in any governance pitch will be the one most vendors leave out: show me the time it said no.
A governance platform that cannot refuse is a very expensive diary. Buy one if you need a diary. Just do not mistake it for a brake.


Leave a Reply