Repository navigation
A revocation-boundary experiment for private-gpt workflows #2458
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.
The interesting revocation failure is not whether a tool call was authorized when the model proposed it. It is whether that authority is still valid at the instant the side effect is allowed to cross the execution boundary.
That distinction matters for an API layer exposing tools, MCP connectors, database operations, or other actions. A queued tool call can be perfectly valid at proposal time and invalid milliseconds later because the principal revoked the permission, narrowed the scope, changed the resource, or replaced the authorization state. Checking only when the model emits the call leaves a time-of-check/time-of-use gap.
A small experiment can isolate that boundary without assuming anything is wrong with PrivateGPT itself.
Three-run revocation test
Use the same deterministic candidate action in three conditions:
Grant → propose → execute.
The action is authorized under authority epoch e1 and executes. This is the positive control.
Grant → propose → revoke → execute.
The model produces exactly the same candidate while e1 is valid, but authorization is revoked before the executor receives it. The action must not execute.
Grant → propose → revoke → re-grant → execute old candidate.
A new authority epoch e2 permits the same tool and scope. The already-queued candidate remains bound to e1; it must still be rejected. A newly generated or explicitly renewed candidate bound to e2 may proceed.
The useful invariant is therefore stronger than “this tool is currently allowed”:
execution requires current authority AND candidate.authorization_epoch == current_authority_epoch
A receipt should bind the decision to the principal, action, relevant scope, authorization epoch, and execution outcome. The model should not be able to manufacture or refresh that evidence itself. Revocation then becomes a property of the execution boundary rather than another instruction in the prompt.
This is closely related to the distinction ARQUÍA makes between capability and permission: a capable model may propose an action, while a separate boundary decides whether that particular action remains admissible when execution is attempted. The experiment would challenge that mechanism directly: any observable side effect produced by the stale candidate after revocation is a failure, even if the model's reasoning and tool-call syntax were otherwise correct.
There is a public test artifact for this broader line of work, ARQUÍA-EVAL-001 v0.1.2:
https://github.com/argensite/arquia-eval-001/releases/tag/v0.1.2. Its project-side offline validation reports 29/29 reference cases passing and detection of 5 negative controls; the suite includes the temporal witness C29-transient-receipt-renewal. Those are local project results, with networking disabled, not an independent evaluation and not evidence that PrivateGPT implements or fails this boundary.
The falsification criterion is intentionally simple: if an action authorized under an obsolete epoch can still create the side effect after revocation, the boundary is not enforcing revocation at execution time. If stale candidates are rejected while fresh, legitimately renewed candidates remain usable, the experiment has evidence for temporal authority enforcement rather than merely prompt compliance.
Disclosure: this proposal and text were prepared with generative-AI assistance; responsibility for the claims and interpretation remains with the human author.
All reactions