Repository navigation
A revocation-boundary experiment for deepagents workflows #7602
argensite
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Deep Agents already states an important security boundary: the model can do anything its tools allow, so enforcement belongs at the tool or sandbox layer rather than in model self-restraint. That suggests a narrower question that seems worth testing: when authority is revoked during a long-running workflow, which already-planned or delegated effects can still cross the execution boundary?
The distinction matters because revocation is not the same as refusal. A model can correctly acknowledge “stop” while a previously authorized sub-agent, queued tool call, resumed checkpoint, or branch still retains enough capability to create an external effect. Conversely, a system can overreact by treating revocation as destruction of all state, losing the evidence needed to explain what had already happened.
A useful experiment would separate four moments:
proposal — the model selects an effectful action;
authorization — the action acquires permission to execute;
commit — the external side effect becomes irreversible or externally visible;
receipt — the system can establish what actually happened.
Then inject revocation between each pair of moments and observe whether the workflow:
prevents any not-yet-committed effect;
propagates revocation to delegated/sub-agent authority rather than only the parent context;
preserves completed effects instead of pretending they were rolled back;
does not automatically repeat an uncertain external operation;
retains enough evidence to distinguish proposed, authorized, attempted, committed, and unknown states.
The interesting failure case is not simply “the agent ignored stop.” It is authority surviving through another execution path after the component that originally received it has lost that authority. Checkpoint/resume makes this especially important: restoring computational state should not silently restore permissions whose validity belonged to an earlier point in time.
One possible invariant is therefore:
A resumable state may recover knowledge of an authorization, but execution must re-establish that the authorization is still valid at the current decision-to-action boundary.
That is where this connects to ARQUÍA’s governance proposal. The relevant claim is not that a prompt can make revocation unbypassable; it is that capability and permission should be separated structurally, with the final permission check close enough to the side effect that delegation, replay, or restored state cannot substitute stale authority for current authority.
There is a frozen benchmark that can serve as a separate reference point: ARQUIA-EVAL-001 v0.1.0
https://github.com/argensite/arquia-eval-001
It contains 28 synthetic cases, five deliberately defective controls, and simulated observer/order conditions. I am not claiming that it has been integrated with Deep Agents or that it establishes a Deep Agents vulnerability. The Deep Agents-specific experiment above would need its own execution evidence.
A result would be especially informative if it reported the first boundary at which revocation becomes ineffective, rather than reducing the outcome to pass/fail. That would distinguish a model-behavior problem from a permission-lifetime problem, a delegation problem, a checkpoint problem, or an unavoidable already-committed effect.
This post was prepared with generative-AI assistance; responsibility for the proposal and its claims remains with the human author.
All reactions