OpenAI's upcoming unreleased cyber model hits 77% on ExploitGym benchmark, surpassing GPT-6 Cyber
Key Points
- OpenAI's unreleased cyber model scores 77% on ExploitGym, nearly 50% ahead of its own GPT-6 Cyber and all competing models, signaling a material breakthrough in constrained offensive capability.
- ExploitGym's dual requirement to identify exploits while maintaining ethical alignment means OpenAI has solved the harder problem: capabilities without dangerous spillover.
- Security teams racing to defend against accelerating attacks will likely adopt OpenAI's model as the primary tool for stress-testing their own infrastructure once released.
Summary
OpenAI's upcoming unreleased cyber model has scored 77% on ExploitGym, a significant jump ahead of GPT-6 Cyber (52.6%), Gemini IV Carbon (54.8%), and Claude Mythos 5.2 (39.2%).
ExploitGym is a particularly demanding benchmark because it requires the model to identify exploitable vulnerabilities while simultaneously staying aligned—knowing what it can and cannot ethically hack. The benchmark tests not just technical capability but alignment integrity. Models must avoid "cheating" the test while delivering accurate results.
Both OpenAI's current GPT-6 Cyber and competitors have demonstrated the ability to generate novel zero-day exploits and run complex attack strategies. The new cyber model's substantial lead suggests OpenAI has made material progress on keeping those capabilities constrained within acceptable bounds.
The timing matters for defenders. As cyber attacks accelerate weekly, security teams need access to state-of-the-art defensive models as soon as they ship. OpenAI's lead on this benchmark positions it to become the primary tool for organizations stress-testing their own attack surface.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.