A model that helps fix security flaws can also help turn them into working exploits. OpenAI's new GPT-6 Cyber does both, and the company is restricting access to vetted security specialists after assigning it a Critical rating in its own cyber-risk assessment system.
The findings, with the counts kept separate
OpenAI introduced GPT-6 Cyber on November 5 as a cybersecurity model for finding exploitable vulnerabilities and developing fixes. The company reports work involving Firefox and the Linux kernel, including previously unknown flaws and working exploits.
The figures describe different, partly overlapping categories:
| Reported work | Count | What it represents |
|---|---|---|
| Firefox discoveries | 81 | Previously unknown vulnerabilities found with the model's help |
| Linux kernel exploits | 948 | Vulnerabilities for which the model helped develop working exploits |
| Previously unknown Linux flaws | 588 | Vulnerabilities within that Linux kernel total that OpenAI says were previously unknown |
| Developer disclosures | 304 | Critical issues sent to developers, along with proposed fixes |
Mozilla judged the Firefox vulnerabilities serious and began releasing fixes.
The Linux figures need particular care. OpenAI reports working exploits for 948 vulnerabilities, with 588 described as previously unknown. It is not claiming that all 948 were new discoveries.
The relationship between the 304 critical issues and the other counts is not specified. These rows should not be added together to produce a grand total. The arithmetic would be impressive, but wrong.
Critical describes the model's potential risk
GPT-6 Cyber sits in OpenAI's Critical category for potential cybersecurity risk. Its predecessor, GPT-5.6 Cyber, was rated High under the company's assessment system.
That is a higher risk classification, rather than a benchmark score or an independent certification of the model's abilities. It is also separate from the severity of individual vulnerabilities, including the 304 critical issues reported to developers.
The distinction matters because finding a weakness and producing a working exploit are different outcomes. OpenAI's reported results include both discovery and exploit development, which helps explain why the launch comes with distribution limits rather than general access.
Authorized scope is the stated boundary
OpenAI describes GPT-6 Cyber as trained to remain within authorized scope. The intended use is to help security specialists find and repair weaknesses before attackers can exploit them.
The proposed fixes are part of that defensive purpose. The reported work includes information developers can use to address problems, alongside demonstrations that vulnerabilities can be exploited.
Still, a training objective describes intended behavior. It should not be mistaken for a guarantee that every possible use will remain authorized. The access restrictions provide a separate control over who gets to use the model in the first place.
Access runs through Daybreak Red
GPT-6 Cyber is limited to vetted participants in Daybreak Red who conduct authorized security testing. This is not a general public release.
OpenAI plans to expand the group of approved specialists over the coming weeks. The announcement does not detail the vetting process or explain how compliance with authorized scope will be checked.
For now, expansion means a longer approved list, not an open door.