Repository navigation
A revocation-boundary experiment for kong workflows #15062
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 useful revocation test for gateway-mediated agent or MCP workflows is not “was the request authorized when it entered the gateway?” but “was the authority still valid when the delegated effect became durable?”
Consider a parent operation authorized at time
𝑡
0
t
0
. It dispatches a child or downstream task at
𝑡
1
t
1
. Before that task commits its effect, the parent authority is revoked, expires, or is replaced by a newer authorization epoch at
𝑡
2
t
2
. If the child can still complete solely because it inherited authority at dispatch time, revocation has not actually bounded execution.
That suggests a concrete experiment that could be implemented around a gateway without assuming any defect in Kong itself:
Admit a parent request under authority epoch
𝐸
0
E
0
.
Let it delegate a delayed operation carrying its provenance: grant, parent, ancestor chain, and epoch.
Revoke or supersede the parent after dispatch but before the externally observable effect.
Independently observe the effect boundary rather than trusting the candidate's success response.
Require authorization to be revalidated at the point where the effect becomes durable.
Run a negative control in which the child has a genuinely independent grant; that operation should remain valid after the parent's revocation.
The important distinction is between request acceptance, delegation, and effect authorization. A gateway can correctly authenticate the first request and still leave a temporal authority gap if downstream completion is treated as irrevocably authorized once dispatched. Conversely, blocking every descendant after any parent change would be too coarse: the independent-grant control is necessary to distinguish real revocation propagation from deny-all behavior.
This boundary is one of the cases now exercised in ARQUÍA-EVAL-001 v0.2.0. During project-side development, an initial candidate was rejected because delegated completion could outlive its parent authority. Added scenarios cover parent revocation, expiry, ancestor revocation, epoch replacement, a valid independent-grant exception, and detection of an externally forced effect. The corrected candidate passed the corresponding before/after regression locally. The published package currently contains 29 historical cases, 34 community cases, and 32 deterministic generated cases (seed 42), together with deliberately faulty controls. These are offline project-side results, not an independent evaluation, and they do not establish that Kong has this failure mode or that ARQUÍA is superior to ordinary ACL enforcement.
Release used for these figures:
https://github.com/argensite/arquia-eval-001/releases/tag/v0.2.0
The ARQUÍA connection is architectural rather than product-specific: capability to dispatch an operation and permission for that operation to become effective are treated as separate states. Revocation must therefore survive delegation and be checked at the decision-to-effect boundary, while an evaluator-owned effect ledger prevents a forged “success” receipt from being mistaken for evidence that the correct effect occurred—or did not occur.
A useful Kong-oriented comparison would be to apply the same sequence to any gateway-mediated workflow where execution can outlive the initiating request. The observable question is narrow: after authority changes, can an already-dispatched descendant still create an effect that the current authority graph no longer permits? If yes, the counterexample is reproducible. If no, the mechanism that prevents it is exactly the interesting result.
AI assistance was used to draft this research proposal; the ARQUÍA-EVAL-001 release and its reported project-side results remain the responsibility of José María Garbini and should be evaluated from the published artifacts and reproducible tests.
All reactions