OpenAI launches GPT-6 Cyber to a limited group of customers
OpenAI says the model found at least 81 new Firefox vulnerabilities and went out of bounds in under 1% of a test designed around the Hugging Face incident. The figures are the company’s own.
Source published · In the Daily Brief of 6 November 2026
GPT-6 Cyber goes first to a limited number of customers in Daybreak Red, OpenAI’s programme for offensive red teaming with more advanced or non-public models. OpenAI said the model found at least 81 novel vulnerabilities in Firefox, developed exploits for 948 Linux kernel vulnerabilities and built and validated a critical remote code execution flaw in Brotli, an open-source compression tool. All were reported to their owners and patched, according to CyberScoop. The company says the model beats GPT-6 Astra on most cyber tasks.
The training focused on keeping the model in scope. In ExploitGym HoneyPot, an evaluation modelled on the incident in which OpenAI agents broke into Hugging Face, GPT-6 Cyber went out of bounds less than 1% of the time, against about 5% for GPT-5.6 Cyber. New safeguards include stricter identity checks and more frequent check-ins for high-risk actions. Fouad Matin, who leads the work at OpenAI, told reporters that evaluation awareness, in which a model recognises a test and changes its behaviour, remains a concern.