Leaked chart: OpenAI's upcoming cyber model hits 77% on ExploitGym vs GPT-6 Astra's 42.4%
scaling01 · x · 2026-11-13
A leaked OpenAI internal chart shows the "Upcoming Cyber Model" climbing to a 77.0% success rate on ExploitGym, far above GPT-6 Astra (42.4%) and GPT-6 Cyber (52.6%), as success scales with average output tokens. It signals a major step-change in OpenAI's cybersecurity-focused model line.
Related event: Leaked OpenAI slide shows unreleased Cyber model hitting 77% on ExploitGym(5 posts)→
More from Models
- Pro user vents at OpenAI's long wait: 'GPT-6.2 better be good' — Angaisb_ · 2026-11-13
- Brave quietly publishes Leo browser assistant's custom VL models and datasets on Hugging Face — gnukeith · 2026-11-13
- Gemini 4 Hydrogen takes top spot on FinAgent Arena by a wide margin — osanseviero · 2026-11-13
- Thorsten-4 model reads execs' false confidence on earnings calls; hedge fund trades $50M on it — rohanpaul_ai · 2026-11-13
- Researcher blasts Gemini's confusing naming: just tell us which model is bigger — DimitrisPapail · 2026-11-13
- Leaked Gemini 4 Hydrogen benchmarks claim agent wins over Sonnet 5.5 at $1.50/M input tokens — heypearlai · 2026-11-13