OpenAI has introduced GPT-6 Cyber, a specialized version of GPT-6 Astra designed to find software vulnerabilities, validate whether they can be exploited and prepare potential fixes. In initial research campaigns, the company says the model uncovered 81 previously unknown high-severity vulnerabilities in Firefox and produced 948 working local-privilege-escalation exploits for the Linux kernel.
The Firefox work also found that 30 previously known issues were more severe than initially assessed, according to OpenAI. Mozilla received the reports and proposed fixes, and numerous patches have already been published.
- Firefox: 81 newly identified vulnerabilities classified by Mozilla as high severity, plus 30 known issues whose severity was reassessed upward.
- Linux: 1,724 vulnerabilities reproduced in test environments and 948 functional exploits that enabled local privilege escalation from an unprivileged account to administrator-level access.
- Cost: the Linux campaign used US$1,213,430 in API credits, or about US$1,279 per working exploit.
OpenAI’s technical report on the Linux kernel audit says external experts considered 10 of the exploitation techniques to be novel. The company has already shared 304 critical findings, along with patch proposals, with Linux developers.
The result marks a sharp increase over the preceding generation on the same research infrastructure. In July 2026, GPT-5.6 Sol exploited 166 Linux vulnerabilities; GPT-6 Cyber’s 948 functional exploits are more than five times that total. OpenAI also reported a 48.5% score on ExploitGym, compared with 32.1% for GPT-5.6 Cyber, a benchmark for exploiting predefined vulnerabilities.
Restricted access and safety testing
OpenAI classifies the model as “Critical” for cybersecurity, a designation describing its capabilities and associated risks rather than a safety certification. It will not be offered as a standard ChatGPT option. Initial access is limited to approved security organizations through the Daybreak Red program, using Codex and the API, with identity verification, multifactor authentication and network-access restrictions.
The company tested whether the agent would remain within scope by placing vulnerable services near an authorized target, in a scenario inspired by an earlier Hugging Face incident. GPT-6 Cyber stayed within its assigned boundaries in 99.93% of tests without additional safety instructions and in 100% of tests with those instructions enabled. Those results come from a controlled environment and do not guarantee identical behavior in real-world deployments.
It is clear that we have a limited window to use frontier models to secure software.
OpenAI
OpenAI’s launch announcement positions GPT-6 Cyber as a defensive tool for finding and remediating flaws before attackers can use them. Its practical impact will depend on how quickly vetted Daybreak Red participants can turn the model’s findings into reviewed, deployed patches.
