Rene Brandel

This talk explores common security vulnerabilities found in AI agents by detailing a process of "hacking" Y Combinator batch AI agents. The core thesis is that agent security extends beyond traditional LLM prompt injection to encompass broader system-level risks, emphasizing the need to treat agents with the same security considerations as human users and to avoid building custom code execution environments.