Autonomous AI agents promise to accelerate and scale security testing, but real-world deployments expose sharp edges that demos rarely show. This talk examines common failure modes, including brittle reasoning, unsafe tool use, environment drift, and misleading test coverage. Drawing from hands-on experiments, we will explore where agents break, why they break, and how security teams can design guardrails that keep automation useful rather than dangerous. Attendees will leave with practical lessons for building, evaluating, and trusting AI-driven security testing systems.
Michał Bazyli is a founding cybersecurity researcher focused on applying AI and machine learning to real-world security problems. He has extensive experience in offensive security, smart contract auditing, and building AI-driven security systems across Web2 and Web3 environments. His current work explores the practical limits, failure modes, and safety challenges of autonomous AI agents used in security testing.
