OpenAI releases GPT-6 Cyber with 48.5% ExploitGym score and 99.93% scope adherence

• OpenAI introduces GPT-6 Cyber, a specialized cybersecurity model that outperforms GPT-5.6 Cyber with a 48.5% score on ExploitGym, up from 32.1%.
• The model demonstrates strong alignment by avoiding out-of-scope exploits in 99.93% of tests without deployment prompts and 100% with prompts.
• Real-world applications include identifying 81 high-severity Firefox vulnerabilities and developing 948 working Linux privilege escalation exploits.
• Classified as Critical under the Preparedness Framework, the model is available via API and Codex for approved Daybreak Red customers.

Performance and Benchmarks

• Achieves 40.3% recall on offensive web penetration testing compared to 32.0% for GPT-6 Astra.
• Completes 95% of tasks on RemoteExploitBench, a new benchmark for remote exploit development, while GPT-6 Astra with Daybreak Blue refused 100%.
• Outperforms GPT-6 Astra on most cyber tasks even when Astra does not refuse the request.

Real-World Security Impact

• Mozilla used the model to find 81 novel high-severity vulnerabilities and validate 30 known issues as more severe than previously reported.
• Researchers reproduced 1,724 Linux vulnerabilities and developed working exploits for 948, more than five times the 166 exploited by GPT-5.6 Sol.
• Security teams at Cisco, Calif, CodeRabbit, and other organizations use the model for vulnerability research and code review.

Safeguards and Access

• Includes cyber-specific misalignment monitoring, Private Safety Processing (PSP), and IP allowlisting for Daybreak Red API access.
• Partners with Dragos to evaluate safeguards against critical infrastructure attacks, including power plants and water treatment facilities.
• Initially available to a select group of approved Daybreak Red customers, with plans to expand access to all Daybreak Red customers soon.