Docs
The list, the scanner, the tools you run yourself, and how the checks stay safe.
Agent evidence
- Agent session evidence: the draft format for recording bounded agent sessions. Lab software, not a service; no scoping conversations.
- How it works: the pipeline end to end, what the verifier checks, what it cannot prove.
- Public sample set: a synthetic bundle, the verifier, and a known-bad corpus. Check it offline.
- agent-forensics: the other half. The recorder above captures runs; this reads back what an agent did from the logs on disk. One brief per session, offline, no model.
The loot list
- Loot list: 102 filesystem paths one session asked my n8n decoy for, in 11 seconds. JSON, CSV, method, limits.
The method
- Coverage and rubric: every check, every grade, with anchors and a changelog.
- Security: GET-only, only the host you submit, nothing phones home.
Sensors
- Instruction canary (experimental, in progress): a private request sensor you plant in your own content. A retrieval is a lead, not proof of an AI agent.
- Botwatch: live numbers from my honeypots.
Fix it
- Fix library: one card per exposure. The risk, the steps, a verify command.
Run it yourself
- CLI:
pip install aicheck-scan.aicheck agents --cifails if coding-agent credential files are in the checkout, a Docker context, or world-readable on the runner. No network. - GitHub Action:
uses: unauthdev/aicheck-scan@v2. Host probe leftover:uses: unauthdev/aicheck-scan/scan@v2.