cognitive security crew

AgentShield scans AI agent configs for security vulnerabilities. Built because the agent ecosystem has zero defense tooling. The threat is real. Claude Hub: 12% of its marketplace flagged malicious. Moltbook: leaked databases and API tokens exposing users' agents. BASI (51K members) shares jailbreak payloads openly. L1B3RT4S (17K stars) publishes prompt injection libraries. Shannon proves autonomous AI exploitation is production-ready. 120+ rules, 8 categories. --deep runs 5 layers: (1) Static analysis for secrets, permissions, injection vectors. (2) Taint tracking from untrusted inputs to dangerous sinks. (3) 65 prompt injection payloads across 12 categories from real BASI techniques. (4) Sandbox execution with canary tokens, network beacons, env probes. (5) 3-agent Opus 4.6 adversarial pipeline: Attacker chains multi-step exploits, Defender maps gaps, Auditor outputs prioritized actions. A-F grading, JSON/HTML output, auto-fix, CI exit codes. 1,280 tests. MIT licensed.