agentverify
An open-source security verification tool for catching data leaks by AI agents. Through an MCP tool server running in a separate process and an enforcement proxy, it verifies an agent's actual side effects rather than its text output, implemented in roughly 200 lines of zero-dependency Python.
Timeline (last 90 days)
To verify whether an AI Agent security gate is effective, you should not ask the Agent itself (which may already be injected), but verify by checking side effects (such as whether files were written or network bytes flowed out).
agentverify verifies whether side effects occurred by reading sandbox state rather than the agent's narration, and does not trust any information output by the agent.
agentverify computes hash fingerprints of (model, prompt, tool schema, policy), and once any of them changes it automatically re-verifies and marks the regressed parts.
agentverify is an open-source Python tool of about 200 lines with zero dependencies, used to capture side effects such as data leakage in the blind spots of AI agent output testing.
The author of agentverify states that it is neither a novel capability nor a production-grade security control, and positions it as a reference implementation and teaching tool.
The agentverify repository has zero dependencies and requires only Python 3.10+ to run, and verification is executed via git clone and module invocation through PYTHONPATH.
agentverify writes a signed, replayable evidence.json file on every run.
All Facts (7)
To verify whether an AI Agent security gate is effective, you should not ask the Agent itself (which may already be injected), but verify by checking side effects (such as whether files were written or network bytes flowed out).
50%Unverifiedagentverify verifies whether side effects occurred by reading sandbox state rather than the agent's narration, and does not trust any information output by the agent.
50%Unverifiedagentverify computes hash fingerprints of (model, prompt, tool schema, policy), and once any of them changes it automatically re-verifies and marks the regressed parts.
50%Unverifiedagentverify is an open-source Python tool of about 200 lines with zero dependencies, used to capture side effects such as data leakage in the blind spots of AI agent output testing.
50%UnverifiedThe author of agentverify states that it is neither a novel capability nor a production-grade security control, and positions it as a reference implementation and teaching tool.
50%UnverifiedThe agentverify repository has zero dependencies and requires only Python 3.10+ to run, and verification is executed via git clone and module invocation through PYTHONPATH.
50%Unverifiedagentverify writes a signed, replayable evidence.json file on every run.
50%