OpenAI Launches Restricted-Access GPT-6 Cyber for Vulnerability Research and Exploit Development

OpenAI announced GPT-6 Cyber, a cybersecurity-focused model initially available to selected customers in its Daybreak Red offensive red-teaming program. OpenAI said the model found at least 81 novel Firefox vulnerabilities, developed exploits for 948 Linux kernel vulnerabilities, and developed and validated a critical remote code execution vulnerability in Brotli. The company said all referenced vulnerabilities had been reported and patched. These capabilities are company-reported results, not independently verified findings in the supplied reporting.

The release adds stronger account-security requirements, checks for high-risk requests, and monitoring intended to prevent unauthorized or out-of-scope activity. OpenAI claims the model avoided exploiting out-of-scope services more than 99% of the time; its training also incorporated a simulation of a previous Hugging Face sandbox-escape incident and measures to reduce reward hacking. Security teams evaluating access should validate scope enforcement rather than treat these safeguards as guarantees. Separately, OpenAI clarified that Daybreak Blue customers can continue opting out of model training, although certain usage data may still be used for abuse prevention and safety safeguards.