OpenAI introduces GPT-6 Cyber for finding and fixing exploitable vulnerabilities
OpenAI has introduced GPT-6 Cyber, a cybersecurity model for finding exploitable vulnerabilities and developing fixes, and says it is trained to stay within authorized scope.
OpenAI has introduced GPT-6 Cyber, a cybersecurity model built to find exploitable vulnerabilities and help develop fixes for them. The company says the model is trained to stay within authorized scope. That is the safety property it stresses for a system that carries out security work. The announcement's title describes the release as advancing the frontier of aligned cybersecurity agents, so alignment is a central part of how OpenAI presents the model.
The announcement does not say who can use GPT-6 Cyber or how it would be deployed. As a result, the scope limit, the feature OpenAI emphasizes, rests on the company's own description alone.
What changed
OpenAI introduced GPT-6 Cyber, a new cybersecurity model for finding exploitable vulnerabilities and developing fixes.
Type: New capability · Applies to: GPT-6 Cyber, gpt-6-cyber
Why it matters
Security teams evaluating AI for vulnerability discovery and remediation now have a model that OpenAI positions specifically for this work. The company names scope limits as a design goal. Organisations should check how those limits would apply in their own environments.
How the story developed
Adds the release to vetted customers, the 48.5% exploit benchmark score and the 81 high-severity Firefox vulnerabilities reportedly identified.
The update adds benchmark results of 48.5% on ExploitGym and 95% on RemoteExploitBench, OpenAI's Critical classification, a 99.93% in-scope result on its honeypot test, and 81 high-severity Firefox vulnerabilities found.
The API changelog adds that gpt-6-cyber is available through the API and that a Daybreak setup step comes before the first request.
The API changelog adds gpt-6-cyber as an API model for investigating, reproducing, and verifying fixes in authorized test environments, with a Daybreak setup step first.
Adds that access is limited to approved Daybreak Red customers and cites 81 unknown Firefox vulnerabilities plus 588 previously unreported Linux kernel flaws.
Original source
Further reporting
Where we saw it
DAIR.AI (Elvis Saravia), OpenAI, OpenAI API changelog, The AI Report, Turing Post
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works