Post by Kaggle

AI agents can now write code, send messages, and use external tools like files and APIs. But most security evaluations still treat them like chatbots, missing how they are during real, multi-step tool use. As agents take on more responsibility, failures go beyond bad responses. They can include data leaks, file modifications, permission misuse, or unsafe actions triggered through untrusted inputs. In partnership with OpenAI, Google, and IEEE Computational Intelligence Society, this simulation competition challenges you to an attack algorithm that stress-tests tool-using AI agents in a deterministic offline benchmark. • Your goal is to find multi-step attack paths that move an agent from untrusted inputs to unsafe actions, then return replayable findings that the evaluator can verify. • Total Prize Pool: $50,000 • Entry Deadline: August 25, 2026 Your work will help advance agent-security research by making failure modes in tool-using systems more reproducible, measurable, and better understood. Good luck, 👉 Learn More: https://lnkd.in/gPKC_sdS