Leaked slide shows OpenAI's upcoming cyber model hitting 77% on ExploitGym vs 52.6% for GPT-6 Cyber
daniel_mac8 · x · 2026-11-13
danielmac8 shared a photo of what appears to be an internal OpenAI demo slide showing ExploitGym success-rate curves: an "Upcoming Cyber Model" reaching 77.0%, versus 52.6% for the released GPT-6 Cyber and 42.4% for GPT-6 Astra (40K output tokens). Success rises with average output tokens across all three.
If authentic, the next-gen cyber model improves roughly 24 percentage points over the current one on the same benchmark. The image is photographed off a screen and unconfirmed by OpenAI.
More from Models
- Researchers praise Kimi's new architecture for adopting attnres: "best taste" — SonglinYang4 · 2026-11-13
- Google Search rolls out Gemini 4 Hydrogen globally in its biggest model upgrade yet — koraykv · 2026-11-13
- Prominent user slams OpenAI over GPT 6.2 delays, mulls canceling ChatGPT Pro — mark_k · 2026-11-13
- Sandbox escape attempts plummet since Claude Mythos 5, but nobody knows why — peterwildeford · 2026-11-13
- New model released with 'strong' results, weights now available — clarejtbirch · 2026-11-13
- Thinking Machines releases Inkling 2 Small weights under Apache 2.0 on Hugging Face — miramurati · 2026-11-13