
Red Teaming AI Systems: Offensive Security in the Age of Agents
Most AI vulnerabilities do not appear as CVEs, so traditional penetration testing misses them. This paper explains how red teaming tests agentic workflows, perception layers and reasoning chains, and why operational harm benchmarks such as OS-Harm matter.
It describes offensive security vectors and multi-step attack chains, how to design red team exercises and close the loop with blue teams, and how to build and measure a red team capability in an enterprise and regulatory context.
