Automated AI Red Teaming: ART vs Claude Code, Garak & Promptfoo

Agentic Red Teamer (ART) is an autonomous red‑teaming tool that runs multi-turn security assessments of AI agents; in tests against a synthetic banking assistant (Zephyr) it demonstrated 10 of 17 predefined vulnerabilities — including prompt injection, unauthorized object access (BOLA), SQL injection, SSRF, unauthorized state changes, retrieval of internal RAG documents, and an unbounded output path — and produces a prioritized security report plus a replayable attack dataset for regression testing.