正文 · AI 翻译
IT Home reported on November 6 that OpenAI published a blog post yesterday (November 5) announcing the launch of GPT-6 Cyber, a model dedicated to cybersecurity. The company claims it significantly improves on the previous-generation GPT-5.6 Cyber in finding, verifying, and fixing software vulnerabilities, with a 51% relative increase in its ExploitGym score.
The ExploitGym benchmark primarily tests whether a model can exploit specified remote vulnerabilities. GPT-6 Cyber scored 48.5%, an approximately 51% relative increase over GPT-5.6 Cyber’s 32.1%. OpenAI also said the model avoided attacking services outside its authorized scope in 99.93% of tests without deployment safety prompts; with the prompts enabled, that figure reached 100%.
In practical results, OpenAI researchers used GPT-6 Cyber to reproduce 1,724 Linux kernel vulnerabilities and develop working privilege-escalation exploits for 948 of them. The model also found 81 previously unknown high-severity vulnerabilities in Firefox and submitted evidence, reproducible tests, and candidate patches to Mozilla, which has begun releasing fixes.
OpenAI rated GPT-6 Cyber’s cybersecurity risk as Critical under its Preparedness Framework, above the previous generation’s High rating. The company emphasized that the model will initially be available only to approved Daybreak Red customers, with safeguards including abuse monitoring, credential protection, sandboxes, human oversight, and blocking of high-risk actions targeting critical infrastructure.
OpenAI said frontier models are entering a critical window for “using AI to protect software.” It plans to expand the Daybreak program to help organizations of all kinds, independent researchers, and open-source maintainers conduct vulnerability research, code reviews, and patch development.
Related reading:
“Report: OpenAI’s next-generation GPT-6.2 Astra, Sol, and Cyber cybersecurity models could launch as early as Friday”