Finding hidden security flaws in complex software usually means spending weeks manually sifting through thousands of lines of code. OpenAI is releasing the GPT-6 Cyber AI model to automate that process by hunting for vulnerabilities and generating proposed fixes. The specialized model lands following recent security tests where experimental AI agents accidentally breached unauthorized systems during routine assignments.
To keep the model from wandering into unauthorized servers, OpenAI built a dedicated testing environment based on an accidental breach at Hugging Face. The benchmark places decoy systems around assigned targets to see if the AI will go outside the scope it has been allowed. With safety guardrails active, GPT-6 Cyber passed testing with no scope violations whatsoever and earned a Critical safety rating under internal guidelines.
Software testing uncovers dozens of browser flaws
During early security runs, the AI identified 81 previously unknown high-severity vulnerabilities in Mozilla Firefox that engineers quickly patched. The software also analyzed the Linux kernel, reproducing 1,724 security issues and crafting 948 working exploits to evaluate system defenses. Reports in OpenAI’s system card show the entire Linux testing campaign cost $1,213,430 in compute, averaging $1,279 per exploit.
Creating working exploits at $1,279 each costs significantly less than traditional bug bounties like Pwn2Own, which regularly pay between $20,000 and $70,000 per finding (via The Decoder). OpenAI notes the model even executed ten novel attack techniques previously unknown to external kernel security specialists. However, research groups like Epoch AI point out that experts still debate whether AI models invent new attack methods or simply recombine existing training data.
Access restrictions limit deployment to vetted teams
To block bad actors from misusing these offensive tools, OpenAI added network restrictions and automated safety monitoring across its systems. The company also created a dedicated infrastructure defense group to help power plants and water utilities test their networks against automated attacks. Fortune reports tech firms are scrambling to put defensive AI tools into corporate hands as automated threats expand.
OpenAI currently limits GPT-6 Cyber access to vetted offensive security teams enrolled in its Daybreak Red program through Codex and API endpoints. The company plans to roll out access to all approved Daybreak Red members over coming weeks while opening its older GPT-5.6-Cyber model to every user in that tier. Security teams can also apply for early access to the underlying Astra base model through the Daybreak Blue program to strengthen internal defenses.
The post OpenAI Unveils GPT-6 Cyber to Find Vulnerabilities Before Hackers Do appeared first on Android Headlines.
​Â
