OpenAI has introduced GPT-6 Cyber, a model designed to find software security flaws, build working exploits and suggest patches. It is available through the API and Codex to a select group of Daybreak Red customers, with access expected to expand gradually to other customers in that program.
The model’s capabilities could help security teams find and repair weaknesses, but they also make it important to control what the model can do. OpenAI says it trained GPT-6 Cyber to stay within its assigned boundaries, including using a honeypot inspired by the Hugging Face incident. The company reports that, without special safety instructions, it avoided boundary violations in 99.93 percent of tests, up from 94.49 percent for its predecessor. With safety instructions enabled, it stayed within bounds in all tests.
How did it perform?
On ExploitGym, a test of whether a model can exploit a vulnerability in a remote system, GPT-6 Cyber scored 48.5 percent, according to OpenAI. GPT-5.6 Cyber scored 32.1 percent; GPT-6 Astra, the standard model, scored 42.4 percent. On a separate test finding flaws in web apps without access to their source code, GPT-6 Cyber had a 40.3 percent hit rate, compared with 32.0 percent for GPT-6 Astra.
OpenAI also says GPT-6 Cyber found 81 previously unknown severe vulnerabilities in Firefox. Mozilla accepted the reports and has shipped fixes. In the Linux kernel, the model reproduced 1,724 vulnerabilities and built working exploits for 948. OpenAI says it reported 304 critical findings to maintainers with suggested fixes. These figures come from the company; independent testing of GPT-6 Cyber has not been published.
Access remains limited
The Linux campaign used automated scout and analyst agents to scan about 5,034 security-relevant locations, roughly seven percent of the kernel codebase. The system narrowed these to 2,308 candidates before confirming the vulnerabilities in isolated virtual machines. OpenAI says the run cost $1,213,430 in computing, or $1,279 per working exploit. That is below the $20,000 to $70,000 historical bounty range it cites for comparable findings, though a research cost and a bounty are different measures.
OpenAI recommends using the model in isolated environments with human oversight. It says its safeguards include checking whether actions match a user’s permissions and monitoring for newly disclosed vulnerabilities. The company is also expanding access to GPT-5.6 Cyber and to GPT-6 Astra for more organizations and researchers.
Why it matters
For software maintainers and security teams, the reported Firefox and Linux findings suggest GPT-6 Cyber could help identify flaws and propose repairs. But access is limited, the performance figures come from OpenAI, and independent testing has not yet been published.
