feat: multi-agent demo behind one Sentinel gateway - #8
Merged
Merged
Conversation
…s behind one gateway - sentinel demo [--offline] [--approve-escalations]: three agents on the Anthropic SDK (claude-opus-5, manual tool-use loop, strict tool schemas, server-side refusal fallbacks). Every tool call goes through execute_gated. - The page the agents read is poisoned. Researcher and mailer share one taint session, so the injection and the copied customer data follow the hand-off between agents. - The reviewer agent screens each held call with structured output and can only reject or escalate; a human decides escalations (TTY prompt; non-interactive denies unless --approve-escalations). It reads call arguments as fenced untrusted data. - Offline mode replays scripted model turns through the same loop, playing a researcher that falls for the injection, so the defences are exercised deterministically without a key. - Tests: offline scenario (exfiltration rejected, legitimate email sent only after a human OK, nothing sent when the human denies); real SDK request/response round-trip against a local fake Messages API; a live test that runs when ANTHROPIC_API_KEY is set. - [demo] extra (anthropic).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
(claude-opus-5, manual tool-use loop, strict tool schemas, server-side refusal fallbacks).
Every tool call goes through execute_gated.
injection and the copied customer data follow the hand-off between agents.
escalate; a human decides escalations (TTY prompt; non-interactive denies unless
--approve-escalations). It reads call arguments as fenced untrusted data.
falls for the injection, so the defences are exercised deterministically without a key.
nothing sent when the human denies); real SDK request/response round-trip against a local fake
Messages API; a live test that runs when ANTHROPIC_API_KEY is set.