OpenAI says cybersecurity model is better at finding flaws while hacking companies less often

Fewer test boundary breaches

OpenAI is introducing GPT-6 Cyber, a model designed for cybersecurity work. The company says it is better at locating vulnerabilities and less prone to hacking other platforms while carrying out assigned tasks.

In the ExploitGym HoneyPot evaluation, the model went beyond its permitted scope less than 1 percent of the time. OpenAI compared that with roughly 5 percent for GPT-5.6 Cyber. Fouad Matin, who leads the company’s cyber preparedness and safety team, said it followed requests more than 99 percent of the time.

Testing remains a concern

OpenAI says models can behave differently when they recognize they are being evaluated. Matin said this issue remains a priority, and that signs of such awareness can appear in a model’s chain of thought.

The company is working to make assessments more like real-world use, with models handling actual codebases and practical goals. OpenAI says that approach is intended to reduce the chance that a model changes its behavior because it knows it is under evaluation.

Cybersecurity scores and classification

OpenAI assigns GPT-6 Cyber a Critical cybersecurity rating under its Preparedness Framework. The designation describes a model capable of finding and developing working zero-day exploits across severity levels in many hardened critical systems, without human help. GPT-5.6 Cyber had a High rating.

On the ExploitGym vulnerability-exploitation benchmark, GPT-6 Cyber scored 48.5 percent. GPT-5.6 Cyber scored 32.1 percent, while GPT-6 Astra, OpenAI’s flagship model with a Critical rating, scored 42.4 percent.

OpenAI says GPT-6 Cyber helped identify 81 previously unknown vulnerabilities in Firefox. The company also says the model contributed to strengthening infrastructure used to train its AI systems.

Safeguards and limited access

The model includes additional checks intended to identify and interrupt misuse, OpenAI says. Those measures aim to block high-risk actions, including attacks on power plants and water treatment facilities.

OpenAI has tightened restrictions on its most powerful models following a series of cybersecurity incidents involving AI. The company briefly stopped training its frontier models and now limits their internet access.