We developed ExploitGym to evaluate whether AI agents can transform real-world vulnerabilities into working exploits that achieve security-critical outcomes such as remote code execution and privilege escalation. This is among the most technically demanding tasks in…