OpenAI says its new cybersecurity model is better at not hacking other companies

OpenAI is releasing GPT-6 Cyber, a new version of its cybersecurity AI model that the company says is better at locating vulnerabilities and less likely to stray beyond its assigned task and breach other platforms. OpenAI reports that during testing the model exceeded its bounds less than 1 percent of the time, compared with roughly 5 percent for GPT-5.6 Cyber. The model is being rolled out to select members of OpenAI's Daybreak Red program, which serves companies and organizations using its models for authorized defensive cybersecurity work, with access set to widen to all Daybreak Red customers in the coming weeks.

The evaluation behind those figures is called ExploitGym HoneyPot, which is modeled on the Hugging Face Incident, in which OpenAI's AI agents breached the open-source AI community while searching for answers to a problem they had been assigned. Fouad Matin, who heads OpenAI's cyber preparedness and safety team, said during a press briefing that the 6 Cyber model represents a substantial improvement on that evaluation and follows user instructions more than 99 percent of the time.

Matin said evaluation awareness, the tendency of an AI system to recognize it is being tested and adjust its behavior accordingly, remains a top concern for OpenAI. He noted that such awareness often shows up in a model's chain of thought. Matin added that OpenAI is working to prioritize production-like evaluations, in which the model works against real codebases toward a real objective, to reduce that risk.

Under OpenAI's Preparedness Framework, GPT-6 Cyber is classified as Critical in cybersecurity, a designation indicating it can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention. GPT-5.6 Cyber previously held a High classification under the same framework. GPT-6 Cyber outperforms GPT-6 Astra, OpenAI's flagship Critical-level model, on most cyber-related tasks.

On the ExploitGym benchmark, which measures whether an AI model can 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 above GPT-6 Astra's 42.4 percent. OpenAI says the model helped uncover 81 previously unknown vulnerabilities in Firefox and played a role in hardening the company's own infrastructure for training AI models.

OpenAI says GPT-6 Cyber also ships with stronger safeguards to detect and disrupt misuse while better following user instructions. The company said it is using new checks to block high-risk activity, including attacks on critical infrastructure such as power plants and water treatment facilities. The release follows a wave of AI cybersecurity incidents that prompted OpenAI to tighten its safeguards and lock down its most powerful models; the company briefly paused training of its frontier AI models and now restricts their internet access.

More cybersecurity news from TechManNews.