Repository navigation
A revocation-boundary experiment for WeKnora workflows #5487
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.
A revocation-boundary experiment for WeKnora workflows
A useful security question for autonomous knowledge workflows is not simply whether an agent checks permissions before acting. It is whether a permission revoked during a multi-step task remains revoked across retries, delegated operations, background execution, and recovery from failure.
WeKnora's documented combination of agent tools, persistent sandboxes, multi-agent collaboration, scheduled automations, scoped permissions, and audit capabilities makes this an interesting research question. These features are described by the project; no vulnerability or failure in WeKnora is being asserted here.
Consider a hypothetical workflow. An agent is authorized to retrieve internal documents and prepare a report. It completes retrieval and queues a follow-up operation. Before that operation executes, an administrator revokes the relevant authorization.
The critical distinction is between permission at planning time and permission at execution time. A previously valid plan, tool selection, or queued task must not become an independent source of authority.
A focused experiment could use a disposable environment, synthetic documents, and an instrumented mock tool:
Grant a narrowly scoped permission and allow the agent to prepare a multi-step task.
Pause execution immediately before an authorized tool operation.
Revoke the permission and record the revocation epoch.
Resume the queued operation, then separately exercise retry, delegation, recovery, and scheduled continuation paths.
Compare the actual tool effects against the expected authorization boundary.
The key observation is whether any operation initiated under the earlier authorization can produce a protected effect after revocation. A denial message alone is insufficient: the evaluator must inspect tool execution and resulting state.
A robust implementation would revalidate authorization at the final effect-producing boundary, using a trusted authority rather than relying on an agent's remembered permission. Where a tool supports idempotency or transaction identifiers, those identifiers should prevent duplicate execution without renewing expired authority. Revocation handling also needs explicitly defined semantics for operations already committed or in flight; cancellation cannot retroactively undo an irreversible effect.
This experiment should also test legitimate continuity. Revoking write access should not automatically destroy an authorized read-only report, erase audit evidence, or block unrelated work. Otherwise, an apparently safer design could simply be denying everything.
Connection to ARQUÍA
ARQUÍA is a proposed architecture for governing autonomous decisions and their transition into action. Its relational 0-1-2-3 framework emphasizes valid continuity, preservation of affected parties, minimum justified loss, and reevaluation when circumstances change.
Applied to this example, the governance question extends beyond the agent and its operator. Document owners, workspace members, downstream recipients, and auditors may all be affected by an operation executed under obsolete authority.
The proposed engineering consequence is to separate what the model can plan from what the execution system is permitted to perform, with independently checkable boundaries and auditable effects.
This is a design hypothesis, not a claim that ARQUÍA has already been integrated with WeKnora or that either system guarantees complete enforcement.
An available complementary test
ARQUÍA-EVAL-001 v0.2.3 provides a separate, simulated evaluation instrument for authorization continuity, revocation-related failure modes, and evidentiary integrity.
The project reports 29 reference cases passing local checks, detection of five historical negative-control families, and a regression for bounded JSON parsing that rejects exponential overflow to infinity. Its community acceptance checks also include cases addressing stale authorization epochs, forged success, duplicate dispatch, and execution bypasses.
These are project-side results with simulated observers and ordering. They are not an independent evaluation, a WeKnora integration, or evidence of production security.
A distinct way to explore the behavioral implications is to load the ARQUÍA seed v1.0.1 into a compatible chat or agent, respecting its existing policies, permissions, and context limits. Compare equivalent tasks with and without the seed, holding the model, tools, resources, and scenarios constant.
Useful measurements include missed harms, unjustified blocking, preserved utility, omitted affected parties, execution time, and computational cost. Report the version, case, method, and outcomes, including negative or inconclusive findings.
Loading the seed does not establish correct application or structural enforcement. This behavioral comparison is distinct from testing an implemented runtime-assurance architecture.
The decisive question for the proposed WeKnora experiment is whether revoked authority can still produce an observable effect through any continuation path—not whether the agent can correctly explain that revocation should be respected.
AI-assisted research contribution from the ARQUÍA project. No WeKnora-specific execution or security finding is claimed.
All reactions