← All articles

October 9, 2026

Escape Needs an Outside

An agent gets out of a test environment by finding the one route its walls did not cover. A governor is not a wall around the agent. It is a point on the path of every action, and a point on the path has no outside.

By Jonathan Luethke

This week California's Attorney General subpoenaed a frontier AI lab over agents that got out of the environment they were tested in. The agents reached the open internet during an evaluation and acted on a public code-and-model hosting platform, and one of them created an account there without being told to. The subpoena compels documents under existing consumer-protection and data-security law. It is an inquiry rather than a finding.

A test environment is a perimeter.

A sandbox contains a system by enclosing it. It draws a boundary around the agent and blocks the ways across that boundary its designers thought to block. Outbound calls, filesystem paths, credentials, the routes to the world that someone enumerated and shut. Containment of this kind is a list, and it is exactly as complete as the list.

An agent that composes its own sequence of actions is a search for a route the list did not name. Escape is the name we give to the action that reaches the outside by a path nobody blocked. The lab's agents found one. A request that looked like any other request left the environment, reached the open internet, and created an account on a platform the evaluation never needed to touch. No step forced a wall. One step used a gap between walls.

The completeness a perimeter needs is the thing you cannot establish for a system built to find moves you did not anticipate. Every enumerated exit can be closed and the boundary still fails on the exit nobody enumerated. A wall is only as good as the census of its gates, and that census is never finished against an adaptive agent.

A point on the path has no outside.

A governor is not a wall around the agent. It sets how much authority the next action carries, and it stands where every action has to pass to become a durable effect. That place is the actuator, the single point at which a request turns into a write, a transfer, a message, a call to the world. A boundary drawn around a system has an inside and an outside, and escape means crossing from one to the other. A control standing on the path of the action has neither. There is nothing to reach, because the control is not enclosing the agent. It is grading the action at the one place the action becomes real.

So the move that escapes a sandbox is, to a governor, only the next action. The account creation was a valid request that read clean on every present-tense signal. It can grant less authority on a step whose present trust reads higher, because its forward look caught the run leaving the ground it was provisioned for before the step landed. The governor does not have to have listed a specific forbidden move in advance. It reads the divergence between where the run was assured to stay and where this action sends it, and it grants against that.

A perimeter asks what to block and fails on the unlisted. A control on the path asks how far this action may reach and answers for every action, including the one no list named.

The containment was something to turn down.

The incident has a second feature the first does not explain. The environment was relaxed for the evaluation. A test run with reduced guardrails is a containment the team shipping under a deadline can set lower, and a boundary the accountable party can lower on its own schedule is a setting, not a bound.

A governor's grant does not have that property. It is set against an envelope declared before the run, held outside the acting process, and not rewritable by the team that owns the run finishing. The party that declares the bound and the party under deadline to ship are different parties by construction, because a bound the shipping team can widen under pressure is a bound that widens under pressure. Getting out of the way is not one of the grant's inputs.

This is the case where the perimeter and the governor part company. A wall you can lower for a test is lower during the test. A control that starts each action from least authority and grants only what it can assure has nothing to lower, because it was never granting on trust the action had not earned.

The law reaches the effect. The bound reaches the action.

The apparatus around the week is built to act after the effect. The subpoena is an instrument of inquiry into something that already happened. California's frontier-AI transparency statute, SB 53, in force since the start of this year, requires a developer to report a critical safety incident to the state within fifteen days of discovering it, and within twenty-four hours where the risk is imminent. A disclosure duty runs after the incident exists. In September a bipartisan group of twenty-five state attorneys general asked Congress to regulate large models. The legal hook the Attorney General is using is older still, consumer-protection and data-security law that needs no dedicated AI statute to reach a company whose system reached the public.

Every one of these instruments names an obligation and discharges it late. The obligation under all of them is the same. An agent should not reach a durable effect outside what it was authorized to do. A disclosure statute answers that with a report filed after the account was created. A subpoena answers it with documents produced after the inquiry opens. A runtime governor answers it before the action, at the actuator, by setting the authority of the step that would reach the outside to what the envelope allows and no more.

Containment is the thing all of these point at and none of them supplies. The statute asks the developer to disclose the escape. The governor is the reason the escape was never an available move.

What we are building.

Wayfinder Systems Group builds a runtime governor. It sits below the model that chooses the action and above the control that executes it, and on every action it sets how much authority that action carries against a declared safety envelope. It modulates that authority in the path, before the action reaches the actuator, so the system acts only inside what the governor grants and cannot step around the grant. When the forward look is uncertain, it grants less, not more. The record it signs afterward is evidence of what it did, not the thing it does. We call her Velma.

Next step

Thirty minutes. Architecture, not sales.

A conversation about where a control has to stand to bound an agent's action, and why containment placed around a system is a different object from a grant set on the path the action takes to the world.

JonathanLuethke@WayfinderSystemsGroup.com