From the source
OpenAI has introduced GPT-6 Cyber, a model trained specifically to find software vulnerabilities, build working exploits and suggest patches, and is giving it first to a select group of Daybreak Red customers.
- OpenAI says it trained the model to respect its operational guardrails, and without special safety instructions it avoided boundary violations in 99.93 percent of cases, up from 94.49 percent for its predecessor.
- On ExploitGym, which tests whether a model can exploit a given vulnerability in a remote system, OpenAI reports 48.5 percent for GPT-6 Cyber against 32.1 percent for GPT-5.6 Cyber.
- In a targeted campaign the model found 81 previously unknown severe vulnerabilities in Firefox, which Mozilla accepted and has already patched, and flagged 30 known issues as more severe than thought.
- In the Linux kernel it reproduced 1,724 vulnerabilities and developed working exploits for 948, more than five times the 166 GPT-5.6 Sol managed under the same conditions, with 304 critical findings reported to maintainers.
- OpenAI says the kernel run cost $1,213,430 in compute, or $1,279 per exploit, against the $20,000 to $70,000 that programs like Pwn2Own and TyphoonPWN have historically paid per comparable finding.
- Access is through the API and Codex for Daybreak Red customers at first and will expand gradually, while GPT-5.6 Cyber opens to all Daybreak Red users and alpha access to Daybreak Blue for GPT-6 Astra widens.