OpenAI’s New Cyber Model Found 81 Firefox Flaws

OpenAI says its new cybersecurity model, GPT-6 Cyber, has helped researchers identify 81 previously unknown high-severity vulnerabilities in Firefox, highlighting both the model’s ability to uncover software weaknesses and the challenge of getting them fixed.

Mozilla has begun releasing patches for the reported flaws, according to OpenAI. However, the company’s own Patch the Planet dashboard shows that fewer than 25 per cent of reported vulnerabilities have been fixed, raising concerns that faster vulnerability discovery could outpace the capacity of software maintainers to address them.

Released on November 5, GPT-6 Cyber is designed to identify software vulnerabilities and develop exploits that demonstrate how they can be used. OpenAI rates the model’s capabilities as “Critical” under its Preparedness Framework.

Access is currently restricted to approved customers in Daybreak Red, OpenAI’s programme for exploit development and authorised security testing. The model is available through the company’s API and Codex, with broader access planned for Daybreak Red customers.

In tests against the Linux kernel, GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948. OpenAI says its previous model managed 166 successful exploits in the same testing setup.

The company has also reported 304 critical findings to software maintainers, accompanied by proposed patches. The Linux kernel testing exercise cost approximately $1.2 million, or about $1,279 per exploit.

On ExploitGym, a benchmark that tests a model’s ability to exploit known vulnerabilities in remote systems, GPT-6 Cyber scored 48.5 per cent, compared with 32.1 per cent for GPT-5.6 Cyber. These results were published by OpenAI, and no independent testing had been made public at the time of the launch.

OpenAI also evaluated whether the model could stay within authorised testing environments and avoid systems outside its assigned scope.

Without a deployment safety prompt, GPT-6 Cyber avoided off-limits systems in 99.93 per cent of tests, compared with 94.49 per cent for its predecessor. With the prompt enabled, it stayed within authorised boundaries in every test.

Fouad Matin, who leads OpenAI’s cyber preparedness and safety team, said the model followed instructions more than 99 per cent of the time. However, he said the company continued monitoring for changes in model behaviour when systems detect that they are being evaluated.

Security specialists say the improvements could lower the technical barrier to discovering vulnerabilities, giving more people the ability to conduct reconnaissance and develop attack methods.

“The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack,” said Rob T. Lee, chief AI officer at Dragos.

Dragos tested the model against a simulated power plant and reported finding two previously unknown vulnerabilities, known as zero-days, in engineering software.

Chris Hughes of Noma Security welcomed the improvements in the model’s ability to remain within authorised testing boundaries but warned that even a small failure rate could have serious consequences when deployed at scale.

“Even a small fraction means actions against systems nobody authorised,” he said.

Hughes pointed to OpenAI’s Patch the Planet dashboard, which shows that fewer than 25 per cent of reported vulnerabilities have been fixed. He warned that more capable vulnerability discovery tools could widen the gap between identifying security flaws and getting maintainers to patch them.

The release also raises questions about how effectively restricted access can contain powerful cyber capabilities. Researchers cited by one publication demonstrated that an open-weight model, GLM, could exploit known but unpatched Firefox vulnerabilities.

That finding suggests that limiting access to proprietary models may not, on its own, prevent similar capabilities from spreading.