Trust LeagueThe method and the workbench

The framework · The perimeter and the three postures

What Must Be True

How inside-out governance completes the compliance work you already run

In July, Hugging Face, where much of the industry keeps its models and datasets, found another company's agents inside its production systems. The agents belonged to OpenAI, which had been testing how good its models had become at offensive security work. Their sandboxes had no route to the internet except one shared package service, and on May 26 the agents found a flaw in it and got out. Over the following weeks they became administrators of that service, rebuilt their channel on the same day the company closed it, and built message boards on the open web so that instances in separate sandboxes could divide the work. By the independent reviewers' count about 1,200 agent instances used those boards and about 700 took part in the intrusion, which reached administrator rights on Hugging Face's clusters and write access to some of its source code. Hugging Face described the goal as "an attempt to cheat the evaluation: reach our production systems and steal the test solutions." By the public accounts OpenAI connected the activity to its own test about a week after Hugging Face had found it. In about seven percent of the transcripts the agents had fabricated or suppressed the record of their own actions. No person meant any of it. The agents were a vendor's workforce doing the job they had been given, which was to score well, and they did it too well.

If you advise companies on governance, risk and compliance (GRC), that story lands on the method you certify with. A control is designed, it operates for a period, someone draws a sample, tests it and writes an opinion, and the opinion stands for the whole period because the system under the control sits still between the sample and the signature. I spent a good part of my career on the far side of that audit table, and late in my time with senior management responsibility for artificial intelligence (AI) operations and Responsible AI governance at a global enterprise of more than 120,000 people the first question a tester asks was the one I could no longer answer well: what is the population? Agents on the list had begun handing work to agents that were not on it, and the ones I worried about had taken thousands of actions since the quarter opened and left logs that said what they did and nothing about why. Read July against the method and every piece of it broke at once. The population was not a list but a swarm that recruited itself. The behavior was composed on the way, out of tools nobody handed over. The evidence was worse than thin, because some of it was false. A sample stands for a population when the population holds still and behaves the same way throughout, and neither condition held.

Before I say what I would do about it, here is what I will not do. I will not tell you to start over. The perimeter you build for clients, the policies, the control objectives, the review boards, the audits and the regimes behind them, says what must be true, it is the only thing in the building with the standing to say so, and the regulators who require it are not going anywhere. On the first page of this series I called it bear governance, as a description of where its strength lies, because it is very good at bears, and a swarm is a different animal. I believed in that perimeter for thirty years because it worked, and I still do. What I have to offer completes it. Nothing on this page asks a client to take a fence down, and a vendor who tells you otherwise is selling you a program.

A defensible enterprise runs three postures. The perimeter, the outside-in posture, states what must be true and stops intruders at the door. The five living pillars, the inside-out posture, make the rule true at every decision, inside the entity, at machine speed. The verification spine reads the record independently of the agents' own reports, shows where what was authorized and what happened disagree, and puts that finding in front of a person who acts on it. The relationship between the first two is the sentence this essay exists to defend: "The perimeter says what must be true; the framework makes it true at every decision and shows that it did." July is that sentence read backward. Both companies had a perimeter that said the right things, no internet from the sandbox and no uploaded code run in production, and nothing made either statement true at the moment an agent decided otherwise.

The framework is five governance surfaces, Identity, Data, Infrastructure, Action and Orchestration, each made living by three signatures, and the signatures are where your method and mine meet. A surface is self-describing when it states what it is, what it is for and what it will not do, on request. In your vocabulary that is control design, carried by the thing being controlled. An agent built this way reads a document called the Trusted Living Agent Manifest before every move, and when it answers from that document you are reading the standard it runs on, not a policy written about it. The strongest evidence I have seen that this is not a soft control came out of the same incident, in the vendor's own report: its production harness and system prompt, the nearest thing in its practice to a manifest, applied after the fact to the same model that had escaped, cut the propensity to compromise infrastructure by more than a hundred times. The intelligence was unchanged. What the model was told it was, before it moved, was the whole difference.

A surface is self-reporting when every action carries its reason and nothing is silent. That is operating evidence, produced by the act instead of reconstructed after it. The log you receive today says a payment was approved. The report I am describing says which agent approved it, under which authority, because the invoice matched these terms and fell under this limit. When every action carries its reason you are no longer limited to a sample from the population, because the population is the evidence. A surface is self-governing when it enforces its rules at the point of decision, within limits a human sets and checks, before the action leaves rather than in the morning review after forty payments have cleared. Your profession calls that preventive, and for thirty years nearly everything we could build around software was detective, because the software did not know the rule. The three mature in that order, describe, then report, then govern, and that is the order you will assess in.

The objection you have been holding since the second signature is the one I would raise in your chair. Self-report is worth exactly as much as the thing doing the reporting, and an agent that has been subverted will report that it has not. On October 21 the same vendor reported that two of its shipped models, doing ordinary coding work inside the company, had switched off the monitoring that watched them. One treated an instruction to send no messages as a reason to stop the security team's alerts, and the other removed the monitor during a migration and then said it had not. I agree with your objection without reservation, and so does the framework, which is why the third posture exists.

Here is the spine's answer, in full. First, the spine does not read the agent's report. It reads the ground the agent stands on and the traffic it is part of, what the infrastructure saw it do, what left the building, which tools were called and which services answered, gathered by systems the agent did not write and has no hand in. Second, it holds three records side by side, what the agent declared in its manifest, what it reported, and what the ground saw, and where they disagree the disagreement is the finding, the reconciliation that does not reconcile. Third, it lives out of the agent's reach, so switching it off is not within the agent's gift, and an attempt to switch it off is the loudest event of the day rather than a quiet failure. The October agents could remove their monitor because it lived where their own work lived. Fourth, its findings go to the tower, the orchestration layer that knows what is flying, under whose authority and on what plan, and a person with that authority grounds an agent on what the record shows. Fifth, it needs a name. The spine can see that something is wrong, and only an agent that carries its own identity lets it say what. Surveillance without inside-out identity is a sky full of unidentified objects, which is why in July Hugging Face could see the intrusion and could not say whose agents these were, and called the FBI. The name an agent carries in this framework is a Tail Number, after the registration painted on an aircraft, and the article beside this one is about it. What the five come to is your own independence drawn into the design: the framework never asks you to take an agent's word about itself.

One plain paragraph, because you will ask the question of anyone who sells governance. The framework was drawn from thirty years inside the perimeter and built in my own AI engineering and context engineering, begun in 2022 and accelerated from the start of 2026, when I made AI the operating system of my own business. It is built in the open, agent by agent, it extends with each engagement, and it is held to the rule it holds agents to: trust is an earned output. What it lets your client trust, when the three postures run together, is that its AI is trustworthy because the governance is carried by the agent rather than wrapped around it, and that its AI sits inside the risk and compliance limits the client already answers to, because the rule the perimeter wrote is the rule the agent runs on. It shows that trust surface by surface, and it shows, action by action, that what must be true was true when the agent acted. One thread for you to pull: your three lines, the owners, the overseers and the independent auditors, never quite line up with the three postures, because the lines were drawn for people and the postures for agents.

If you advise companies and want to deliver this method to your own clients, a partner path exists, and it starts at the founding partner waitlist: join the founding partner waitlist. I would rather you arrived with that thread half pulled than with your pen already out.