AI masters cyberattacks: from CTF flag to working exploit

Dor Cohen published an analysis showing that AI models improved their performance on InterCode-CTF from 35% to 95% over nine months, while on ExploitBench, the share of models capable of turning bugs into code execution grew from ~2% in May 2026 to 73% for Claude Opus 5.5 by September 2026.

Compiled by Sergey KostenchukPublished 2026-11-08Updated 2026-11-08

2026-11-08 Safety OpenAI

🛡 AI has taken hacking to working exploits

Dor Cohen analyzed how quickly models are learning to hack: they improved their performance on InterCode-CTF from 35% to 95% in nine months. OpenAI closed CVE-Bench (GPT-5.3-Codex: 90% of tasks, February 2026), and Anthropic closed CyberGym (83% for Claude Mythos Preview, April 2026). The current frontier is ExploitBench: the share of models capable of turning bugs into code execution grew from ~2% in May 2026 to 73% for Claude Opus 5.5 by September 2026.

🌍 According to the author, public hacking benchmarks saturate within 9–11 months, and vendors are already closing their own. The industry needs new, more complex metrics; defenses are working—successful exploits have dropped from 157 to 45 (ExploitGym).

👤 The boundary is now visible: models are not just finding bugs but turning them into code execution—for example, in V8/Chrome. Progress can be tracked via live benchmarks.

Source 1: https://dorcohen.dev/posts/what-can-ai-actually-hack/ Source 2: https://exploitbench.ai/

Sources