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

OpenAI is launching GPT-6 Cyber, the latest version of its cybersecurity AI model, which it says is better at finding vulnerabilities and less likely to go rogue and hack other platforms in the process. OpenAI says that during testing, GPT-6 Cyber went out of bounds less than 1 percent of the time, down from around 5 percent with GPT-5.6 Cyber.

The evaluation, called ExploitGym HoneyPot, is modeled after the Hugging Face Incident, in which OpenAI’s AI agents breached the open-source AI community to find answers to a problem they were tasked with solving. “With the 6 Cyber model, we were able to make huge improvements on this evaluation,” Fouad Matin, the head of OpenAI’s cyber preparedness and safety team, said during a press briefing. “It’s over 99 percent of the time doing what you ask.”

But evaluation awareness, or when an AI system recognizes that it’s being tested and changes its behavior in response, remains a concern. “Evaluation awareness is super top of mind for us,” Matin said. “Often when we do see evaluation awareness, you’ll see it in the model’s chain of thought.” Matin added that OpenAI is trying to prioritize production-like evaluations, where the model is interacting with real codebases with a real objective to lessen the risk of awareness.

Following a wave of AI cybersecurity incidents, OpenAI has raced to strengthen its safeguards and lock down its most powerful models. The company briefly paused training of its frontier AI models, and it now restricts their access to the internet.

GPT-6 Cyber is classified as Critical in cybersecurity on OpenAI’s Preparedness Framework, which indicates that 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 was previously classified as High under the framework.

GPT-6 Cyber outperforms GPT-6 Astra, OpenAI’s flagship Critical-level model, on most cyber-related tasks. In ExploitGym, an AI benchmark that measures whether an AI model can exploit vulnerabilities, GPT-6 Cyber scored 48.5 percent, about 16 points higher than GPT-5.6 Cyber’s 32.1 percent and a few points higher than GPT-6 Astra’s 42.4 percent score. The model helped find 81 previously unknown vulnerabilities in Firefox, and it played a role in hardening OpenAI’s own infrastructure for training AI models, OpenAI says.

Additionally, GPT-6 Cyber comes with stronger safeguards “to detect and disrupt misuse” while better adhering to user instructions. OpenAI says it’s using new checks to block “high-risk activity,” including attacks on critical infrastructure like power plants and water treatment facilities.

GPT-6 Cyber is rolling out to select members of OpenAI’s Daybreak Red program, which is available to companies and organizations using its models for “authorized defensive cybersecurity work.” It’s expanding GPT-6 Cyber to all Daybreak Red customers in the coming weeks.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.