Leaked OpenAI slide shows unreleased Cyber model hitting 77% on ExploitGym

On November 13, OpenAI President Greg Brockman publicly discussed a cybersecurity model the company is developing, calling it a "step-change leap" with ongoing progress. Soon after, screenshots—reportedly from inside OpenAI or a live talk—surfaced showing an unreleased model labeled "Upcoming Cyber Model" delivering standout performance on the exploit benchmark ExploitGym: success rates climbing to 77.0% as the output token budget grows, versus about 42.4% for GPT-6 Astra and roughly 52% for GPT-6 Cyber, putting the new model at nearly double the comparison models.

Confirmed

• Brockman himself confirmed OpenAI is developing a cyber-focused model and described it as a step-change leap
• Multiple posts (danielmac8, Kyrannio, etc.) shared the same demo screenshot with consistent numbers: Upcoming Cyber Model at 77.0%, GPT-6 Astra at 42.4%, GPT-6 Cyber around 52%
• The screenshot shows success rates rising with output token count, forming a scaling curve

Unconfirmed

• The screenshot is "alleged," taken from a livestream or internal slides; OpenAI has not officially released this benchmark data
• The Cyber model's exact name, version, and release date remain undisclosed

Why it matters

• 77.0% nearly doubles the 30%-50% range typical of comparable models; if real, it marks a major leap in offensive/defensive security
• ExploitGym measures exploitation capability, and such large gains have sparked discussion about potential misuse for cyberattacks and related safety risks

2026-11-13 ~ 2026-11-13 · 5 related posts

• [source] Greg Brockman: OpenAI's internal cyber model is 'a step function further', scoring 77.0% on ExploitGym — ChrisGPT · 2026-11-13
• [source] Leaked chart: OpenAI's upcoming cyber model hits 77% on ExploitGym vs GPT-6 Astra's 42.4% — scaling01 · 2026-11-13
• Livestream Leak: Upcoming Cyber Model Hits 77% on ExploitGym vs Rivals' 30-50% — Kyrannio · 2026-11-13

2 near-duplicate retellings: daniel_mac8 · daniel_mac8