OpenAI releases GPT-6 Cyber, a model that found 81 Firefox vulnerabilities
A specialized model for authorized vulnerability research and exploitation is available only to vetted Daybreak Red customers. OpenAI rates its cybersecurity capability as Critical, the highest tier in its safety framework.
From vulnerability hunt to working exploit
GPT-6 Cyber is a specialized variant of GPT-6 Astra built for security work: reproducing crashes, developing exploits, searching large codebases and testing systems remotely. These are tasks that general models often refuse or handle only partially.
The numbers are unusually concrete. OpenAI says the model helped find 81 previously unknown high-severity vulnerabilities in Firefox and turned 948 of 2,308 candidate Linux kernel privilege-escalation flaws into working local root exploits. Mozilla has already patched many of the Firefox findings.
On the ExploitGym benchmark it scores 48.5%, up from 32.1% for GPT-5.6 Cyber. In a black-box web penetration test it reaches 40.3% vulnerability recall against 32.0% for Astra, and completes 95% of RemoteExploitBench tasks. In Irregular's external test it solved 90 of 288 FrontierCyber challenges, while Astra solved 96 and none of the 15 hardest tasks fell.
Access only after verification
The model is not public. It is available through Codex and the API to approved Daybreak Red customers, with identity checks, single sign-on, MFA, managed devices and API IP allowlisting by default. Broader access is promised later without a date.
OpenAI classifies GPT-6 Cyber as Critical for cybersecurity, High for biological and chemical capability and below High for AI self-improvement. The system card reports a lower stopping rate than Astra (73.00% vs 78.38%) and incomplete blocker reporting in 20.29% of responses. In a honeypot test, out-of-scope exploitation occurred in 0.07% of cases without safety instructions and 0% with them.
OpenAI did not publish pricing. Most developers therefore cannot compare cost or throughput with their own security workload yet.
What it means beyond the security team
The same week Google expanded access to Gemini 4 Carbon and Tencent presented a security-trained open model. Frontier labs now treat cyber capability as a product line, not a side effect.
For defenders this is a real advantage: a model that finds a flaw, verifies exploitability and proposes a patch compresses the time between disclosure and fix. For everyone else it shortens the window in which unpatched software is safe.
Sources
More news
Google releases Gemini 4 Carbon: a frontier model for cybersecurity
Google's strongest model leads in finding vulnerabilities and fixing them. Access is initially limited to a handful of cyber defenders — a public release will follow later.
Read articleTencent unveils Hy4: an open model trained only on defensive cybersecurity
The Chinese giant showed a model with vision, computer use and stronger agentic skills. It excluded exploitation material from training and promises weights in the coming days.
Read article