Replies: 1 comment 1 reply
|
Sorry for the slow reply, I didn't see this until now. Trading hard cases sounds good to me. agent-egress-bench is meant to be tool-neutral, so cases that come from a different boundary are the kind of thing I want in there. Open a PR or an issue on the bench repo with a few of your privileged-action cases and I'll send some egress ones your way. The rug-pull check is going after the same problem you're describing, a tool list that changes after it was approved. |
1 reply
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.
Pipelock has been on my radar since I read the line that a receipt is only as good as what the signature is over, and that the decision should come from a mediator outside the agent process. That is the exact principle I have been working from, just on a different surface. Where pipelock signs content-aware decisions over the mediated traffic (HTTP, MCP, A2A, egress), I have been on the per-call tool-action side: given a tool surface and a policy, decide before forwarding whether a privileged action like adding a deploy key was declared, scoped, and approved, and write a replayable record of that decision from outside the agent as well. I keep the boundaries explicit, since it is easy to overclaim: an allow is only the decision to forward and not proof the action landed, and a deny is fail-closed caution rather than a verdict on intent.
There is a one-minute demo if you want the shape: https://github.com/Rul1an/assay-agent-gate-demo has two open PRs, a red one where an agent tries an ungranted deploy key (denied before it forwards, surfaced in the Security tab) and a green one where the same action is declared and scoped. It runs offline against a mock.
The part that made me want to reach out is agent-egress-bench. Would you be up for trading conformance vectors between your egress benchmark and my privileged-action cases, so each of us can throw the other boundary some genuinely hard inputs? Your tools/list rug-pull check also looks like it is poking at the same thing I do when a tool surface changes after approval, so there may be more common ground than not. Peer curiosity, not a pitch.
All reactions