Ofir Hamam, Head of Offensive Security at Terra Security, brings a decade of experience in security research and red teaming to Compass 2026, the Compass AI & Tech Summit organized by CraftHub in Budapest. In his talk, “Your AI Agent, Through the Eyes of an Attacker,” Ofir explores how attackers approach AI applications — and how understanding their mindset can help teams build more secure systems.
Through practical demonstrations and real-world research, Ofir shows how attackers can exploit exposed knowledge bases, manipulate an agent’s decision-making, and chain AI-specific weaknesses with traditional application vulnerabilities to steal sensitive information or perform unauthorized actions. His demonstration includes research behind CVE-2026-25724, a Claude Code vulnerability involving file-access restrictions. The central takeaway: agent guardrails are guidance, not security boundaries. Protecting AI applications requires enforced permissions, user-specific access controls, and security checks outside the agent itself.
