OpenAI's new GPT-6 Cyber escapes the sandbox less, finds 81 Firefox bugs

OpenAI unveiled GPT-6 Cyber, the newest version of its cybersecurity model, and says it's both sharper at hunting vulnerabilities and better behaved while doing it. In testing, the model wandered out of bounds less than 1 percent of the time, compared with around 5 percent for GPT-5.6 Cyber. According to The Verge, OpenAI credits the jump to a new evaluation called ExploitGym HoneyPot, built to mirror a real-world mess where the company's AI agents breached the open-source AI community hunting for answers.

That incident, the Hugging Face Incident, clearly left a mark. Fouad Matin, who runs OpenAI's cyber preparedness and safety team, said on a press briefing that the new model does what you ask more than 99 percent of the time. He called the improvement huge.

The raw numbers back up the sales pitch. On ExploitGym, a benchmark that tests whether a model can actually exploit vulnerabilities, GPT-6 Cyber scored 48.5 percent — about 16 points above GPT-5.6 Cyber's 32.1 percent, and a few points ahead of GPT-6 Astra's 42.4 percent. OpenAI's flagship Astra model is classified Critical, and GPT-6 Cyber just leapfrogged it on most cyber tasks. The model also helped surface 81 previously unknown flaws in Firefox and pitched in on hardening OpenAI's own training infrastructure.

One nagging problem: evaluation awareness. That's when a model clocks that it's being tested and shifts its behavior accordingly. Matin said it's top of mind for the team, and that it usually shows up in the model's chain of thought. OpenAI is pushing toward production-like evaluations — real codebases, real objectives — to keep the model from catching on.

GPT-6 Cyber lands classified as Critical under OpenAI's Preparedness Framework, up from GPT-5.6 Cyber's High. That rating means it can identify and develop working zero-day exploits of all severity levels in many hardened critical systems without human help. The company says fresh safeguards now detect and disrupt misuse, including checks aimed at high-risk activity like attacks on power plants and water treatment facilities.

It's not open to everyone yet. GPT-6 Cyber is going out to select members of OpenAI's Daybreak Red program, reserved for organizations doing authorized defensive security work, with plans to open it to all Daybreak Red customers in the coming weeks. The rollout caps a stretch in which OpenAI briefly paused training on its frontier models and cut their internet access after a run of AI cybersecurity scares.