CyberGym benchmark reveals massive AI hacking leap

CyberGym is graduating from a research project to the definitive metric for agentic security capabilities, effectively becoming the "SWE-bench" of offensive research.

• –Frontier models have jumped from ~12% to over 80% success in less than a year, signaling a breakthrough in long-horizon reasoning and codebase navigation.
• –The discovery of 35 zero-days in major libraries like OpenSSL proves that AI agents can now outperform traditional fuzzing and human review in specific contexts.
• –High performance on CyberGym was reportedly a key factor in Anthropic's decision to gate "Mythos," illustrating how benchmarks are now driving safety policy.
• –The shift to execution-based eval (requiring a working PoC) prevents "data contamination" leaks that plague static security benchmarks.
• –Microsoft’s MDASH now leads the leaderboard at 88.45%, demonstrating the efficacy of multi-agent architectures in complex security tasks.