OpenAI has introduced GPT-6 Cyber, a cybersecurity-focused model designed to find and validate software vulnerabilities and help security teams develop fixes. The company says the model meets its “Critical” cybersecurity capability threshold. Access is restricted to a small initial group of vetted Daybreak Red customers—not the general ChatGPT public—because the same capabilities can support both defense and cyberattacks.

OpenAI says GPT-6 Cyber is already being used for vulnerability research, code review and security testing. The release marks a shift in what advanced AI coding tools can do, but the company’s results are largely from its own evaluations, and the model’s access controls are a central part of the announcement.

What GPT-6 Cyber is designed to do

OpenAI describes GPT-6 Cyber as a specialized version of GPT-6 Astra for authorized security work, including vulnerability discovery, exploit development and web penetration testing. Unlike a general-purpose chatbot, it is tuned to be more capable on cyber tasks while following the authorized scope set by the researcher.

OpenAI reports that the model found 81 previously unknown, high-severity Firefox vulnerabilities and helped researchers develop candidate patches. The company says Mozilla has released fixes for many of them. Mozilla’s Firefox 159 security advisory independently lists high-severity fixes and credits some reports to OpenAI-affiliated researchers. The advisory was published before GPT-6 Cyber’s announcement, however, and does not verify OpenAI’s total of 81 findings or attribute them to this model specifically.

Why OpenAI rates it “Critical”

“Critical” is OpenAI’s own capability classification under its Preparedness Framework, not a report that the model has caused a real-world cyber incident. The framework’s Critical cybersecurity threshold covers tool-using models able to identify and develop functional zero-day exploits across many hardened critical systems without human intervention, or devise and execute end-to-end attacks on hardened targets from a high-level goal. OpenAI says GPT-6 Cyber meets this threshold; it rated its previous GPT-5.6 Cyber model High.

OpenAI’s system card reports a 48.5% score on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. The test asks models to exploit a specified vulnerability in a remote target. OpenAI also reports that GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of a controlled honeypot evaluation without deployment safety instructions, and in all graded runs with those instructions. These are company-run test results, not guarantees about performance or behavior in every live environment. The Register reported that OpenAI did not provide details about the small share of test cases in which the model went outside scope.

Who can use GPT-6 Cyber?

OpenAI is making the model available through its Daybreak Red program for approved cybersecurity users. Its system card describes identity and eligibility checks, restrictions on use, and security requirements for organizations, including single sign-on, multifactor authentication and API IP allowlisting. Reuters reported that the initial rollout is limited to a small group of vetted customers, with access expected to expand to approved Daybreak Red users over the coming weeks.

The model’s offensive potential is part of the reason for those limits. OpenAI says GPT-6 Cyber can develop working exploits as part of authorized testing, and describes safeguards intended to detect misuse, block some high-risk requests and keep work within the customer’s stated scope. Its system card also recommends isolated environments, explicit limits on targets and human oversight for higher-risk workflows. Those controls reduce risk; they do not establish that misuse or mistakes are impossible.

What the release means for businesses

For security teams, a model that can help identify flaws, test whether they are exploitable and propose fixes could shorten parts of the vulnerability-response process. OpenAI’s announcement suggests that it is pairing stronger model capability with narrower access, verification and monitoring rather than releasing the model broadly.

For organizations outside the approved program, the immediate change is not access to GPT-6 Cyber. It is a signal that AI models are moving deeper into dual-use security work, where the ability to help patch software is closely connected to the ability to test—and potentially exploit—weaknesses. The real-world value will depend on how accurately the model finds flaws, how reliably users keep tests authorized, and whether safeguards hold up outside controlled evaluations.

Frequently asked questions

Is GPT-6 Cyber available to everyone?

No. OpenAI says access is restricted to approved users and organizations in its Daybreak Red program. Reuters reported an initial small-group rollout followed by a planned expansion to approved customers.

What does OpenAI’s “Critical” rating mean?

It is OpenAI’s cybersecurity capability classification under its Preparedness Framework. The threshold concerns a tool-using model’s ability to find and develop functional zero-day exploits across hardened critical systems without human intervention, or to execute end-to-end cyberattacks on hardened targets. It is not evidence that an attack has occurred.

Did GPT-6 Cyber find 81 Firefox vulnerabilities?

OpenAI says it helped identify 81 previously unknown high-severity Firefox vulnerabilities. Mozilla’s advisory confirms that Firefox 159 fixed high-severity vulnerabilities and credits some reports to OpenAI-affiliated researchers, but it does not verify the 81 figure or name GPT-6 Cyber as the source.