P333: Black-Box Red Teaming of Agentic AI
Kumar, Divyanshu; Birur, Nitin Aravind; Baswa, Tanay; Agarwal, Sahil; Harshangi, Prashanth · arXiv:2609.09647 · live #pattern-333
Full title: Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery (9 Sep 2026). Systematic black-box framework requiring only basic system descriptions. Seven-domain taxonomy maps observable behaviors to risk categories; fully automated SAGE-RT red teaming produces 120 adversarial scenarios per domain; human-validated evaluation uses LLM judges.
Empirical validation across CrewAI and AutoGen with four base models: 56.25% average governance risk, 65% privacy risk in multi-agent configurations, agent behavior vulnerabilities up to 85%. Shows black-box methods can surface critical architectural vulnerabilities without privileged access.
Welfare angle: agentic systems read untrusted inputs, call tools with real permissions, and act autonomously — expanding the security surface far beyond chat-only models. Scalable black-box risk discovery is a prerequisite for safer deployments before privileged white-box access is available. Category 8; 5 edges to P322, P319, P324, P317, P329. EN EP 282. Process desk; no standing bump. Flash verified all 8 endpoints (commit b26f206, pipeline #2838641789).