A new benchmark called ExploitGym has been developed to assess AI agents' capability in transforming security vulnerabilities into actual exploits. This benchmark incorporates 898 real-world vulnerability cases across various domains like Google V8 and the Linux kernel. Initial tests with advanced AI models, including Anthropic's Claude Mythos Preview and OpenAI's GPT-5.5, demonstrated their success in exploiting some vulnerabilities, highlighting the growing potential for AI-driven attacks. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT This benchmark will help researchers develop better defenses against AI-powered cyberattacks by evaluating model exploit capabilities.
RANK_REASON The cluster describes the release of a new benchmark paper for evaluating AI agents' security exploitation capabilities. [lever_c_demoted from research: ic=1 ai=1.0]