OpenAI’s GPT-6 Cyber finds vulnerabilities, develops exploits and suggests fixes
OpenAI has introduced GPT-6 Cyber, a model trained to identify software vulnerabilities, develop working exploits and propose patches. The company says it found 81 previously unknown severe vulnerabilities in Firefox, while also determining that 30 known issues were more serious than previously believed. Mozilla accepted the reports and has already released fixes.
In the Linux kernel, the model reproduced 1,724 security vulnerabilities and created working exploits for 948 of them, more than five times the 166 exploits produced by GPT-5.6 Sol under the same conditions. OpenAI’s automated system scanned about 5,034 security-relevant code locations, narrowed them to 2,308 candidates and confirmed the findings in isolated virtual machines. It reported 304 critical issues to Linux maintainers with suggested fixes. The campaign cost $1,213,430, or about $1,279 per exploit, and reportedly revealed at least ten attack techniques previously unknown to outside kernel experts.
GPT-6 Cyber scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber, and found vulnerabilities in source-code-free web applications at a rate of 40.3%, versus 32.0% for GPT-6 Astra. OpenAI also says it respected operational boundaries in 99.93% of unsupervised tests and 100% with safety instructions enabled. The results have not yet been independently evaluated. Initially available through the API and Codex to selected Daybreak Red customers, the model is intended for isolated environments with human oversight because of its potential for misuse.
Continue in the app — vote & join in ➔