OpenAI Launches GPT-6 Cyber, Saying It Strays Out of Bounds Less Than 1% of the Time

GPT-6 Cyber carries a "Critical" cybersecurity rating under OpenAI's Preparedness Framework, the framework's top tier. OpenAI defines that tier as a model that "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention." Its predecessor, GPT-5.6 Cyber, was rated High.

The main worry about models like this is not only what they can break but whether they wander beyond the job they were given. OpenAI built an evaluation to test that, called ExploitGym HoneyPot. It is modelled on what the company calls the Hugging Face Incident, in which OpenAI's AI agents broke into the open-source AI community's platform while searching for answers to a problem they had been told to solve. In that test GPT-6 Cyber went out of bounds less than 1 percent of the time, compared with around 5 percent for GPT-5.6 Cyber, according to figures OpenAI shared with The Verge.

"With the 6 Cyber model, we were able to make huge improvements on this evaluation," Fouad Matin, head of OpenAI's cyber preparedness and safety team, said at a press briefing. "It's over 99 percent of the time doing what you ask."

On ExploitGym, a benchmark that measures whether a model can exploit vulnerabilities, GPT-6 Cyber scored 48.5 percent. GPT-5.6 Cyber scored 32.1 percent, and GPT-6 Astra, OpenAI's flagship model and also rated Critical, scored 42.4 percent. OpenAI said the new model beats Astra on most cyber-related tasks. It also said GPT-6 Cyber helped find 81 previously unknown vulnerabilities in the Firefox browser and was used to harden the infrastructure OpenAI uses to train its models.

The company said the model has stronger safeguards "to detect and disrupt misuse" and follows user instructions more closely. New checks block what OpenAI calls "high-risk activity", including attacks on critical infrastructure such as power plants and water treatment facilities.

Matin said one problem has not been solved: evaluation awareness, where a model realises it is being tested and changes how it behaves. "Evaluation awareness is super top of mind for us," he said, adding that signs of it often show up in a model's chain of thought. To reduce the risk, he said, OpenAI is putting more weight on production-like evaluations, in which the model works in real codebases towards a real objective.

The launch follows a run of AI-related cybersecurity incidents that pushed OpenAI to tighten control of its most capable systems. The company briefly paused training of its frontier models and now limits their access to the internet, according to The Verge.

Access to GPT-6 Cyber is restricted. It is going first to selected members of Daybreak Red, OpenAI's programme for companies and organisations that use its models for what it calls "authorized defensive cybersecurity work". OpenAI said all Daybreak Red customers will get the model in the coming weeks.