ChatGPT Cyber is an OpenAI model for cybersecurity research, available to the highest-tier users in its Daybreak Red program. OpenAI says it can help identify software vulnerabilities and prepare fixes, but its reported findings and safety results still require independent evaluation.
Key takeaways
- OpenAI has limited ChatGPT Cyber access to the highest tier of its Daybreak Red cybersecurity research program.
- OpenAI said the model prepared patches for 948 potential Linux kernel issues and identified 81 novel Firefox vulnerabilities.
- The reported findings are not proof that every flagged issue is exploitable or every proposed patch is safe to deploy.
- The Verge reported that the model went out of bounds in less than 1% of testing, compared with around 5% for GPT-5.6 Cyber.
- OpenAI has said it wants to expand Daybreak access, but any wider rollout will depend on access controls and safeguards.
OpenAI has introduced a cybersecurity-focused model for selected users in Daybreak Red, its program for advanced cybersecurity research. The company says the model can help find software vulnerabilities and prepare fixes. Access is limited to the program’s highest tier, placing the release within a controlled research initiative rather than making it broadly available.
The launch highlights a central tension in AI cybersecurity: models that help defenders identify and repair weaknesses may also be useful for offensive work. OpenAI’s reported research results and its safety claims offer an early view of how it is approaching that challenge, though neither replaces independent evaluation.
What OpenAI announced and who can access ChatGPT Cyber
Daybreak Red provides tools for offensive cybersecurity research, including vulnerability discovery and development. OpenAI has made ChatGPT Cyber available to the program’s highest-tier users, rather than opening access to all customers. That restriction gives the company a way to limit use while assessing the model’s capabilities and risks.
Coverage differs on the model’s name. TechCrunch calls it ChatGPT Cyber, while The Verge reports on GPT-6 Cyber. The available reporting does not establish whether those names are interchangeable, so they should not be treated as confirmed aliases. Both accounts describe a new OpenAI model focused on cybersecurity.
Reported Linux and Firefox vulnerability findings
OpenAI said the model scanned Linux kernel code and prepared patches for 948 potential issues. It also reported that the model identified 81 novel vulnerabilities in Firefox. These figures suggest potential value for teams responsible for large, complex codebases, where reviewing code and validating suspected flaws can take substantial effort.
The results are company-reported findings, not a guarantee that every flagged issue is an exploitable vulnerability or that every proposed patch is safe to deploy. Security researchers still need to verify findings, assess their severity, and test fixes. In practice, AI vulnerability research is most useful as part of that human-led process - helping direct attention, not replacing expert review.
Safety claims, the GPT-6 Cyber naming, and what comes next
The Verge reported that GPT-6 Cyber outperformed GPT-6 Astra on cybersecurity benchmarks. It also said the model went out of bounds in less than 1% of testing, compared with around 5% for GPT-5.6 Cyber. Those figures are reported test results; they do not by themselves show how the model will behave across every real-world setting or user request.
The Verge also reported that GPT-6 Cyber received a Critical cybersecurity classification under OpenAI’s Preparedness Framework. The classification reflects the model’s potential to assist with advanced cyber capabilities, making safeguards and restricted access important parts of the release. OpenAI’s reported testing suggests improved boundary-keeping, but the model’s capabilities and the consequences of misuse make ongoing scrutiny essential.
OpenAI has said it wants to expand Daybreak to organizations of different sizes, independent researchers, and open-source maintainers. Any broader rollout will depend on how access and safeguards develop. For now, the launch combines promising vulnerability-research claims with restricted availability and safety measures that will need continued evaluation.
Follow our AI and cybersecurity coverage for updates on Daybreak Red access, independent evaluations, and OpenAI’s plans to broaden the program.
By the numbers
OpenAI said the model prepared patches for 948 potential Linux kernel issues.
This is a company-reported result in the article; it does not establish that every flagged issue is exploitable or every patch is safe to deploy.
OpenAI reported that the model identified 81 novel vulnerabilities in Firefox.
This figure is attributed to OpenAI's reported research and requires independent validation of the findings.
The model went out of bounds in less than 1% of testing, compared with around 5% for GPT-5.6 Cyber.
The Verge reported these test results. They do not establish how the model will behave across all real-world settings or requests.
Frequently asked questions
What is ChatGPT Cyber?
ChatGPT Cyber is an OpenAI model designed to support cybersecurity research. The company says it can help find software vulnerabilities and prepare fixes. Its reported findings should be verified by security researchers before teams act on them.
Who can access ChatGPT Cyber?
Only the highest-tier users in OpenAI's Daybreak Red program can access ChatGPT Cyber, according to the article. Daybreak Red provides tools for advanced cybersecurity research. OpenAI has said it wants to expand access to organizations, independent researchers, and open-source maintainers.
What vulnerabilities did ChatGPT Cyber reportedly find?
OpenAI said the model prepared patches for 948 potential Linux kernel issues and identified 81 novel Firefox vulnerabilities. These are company-reported results. Researchers still need to validate the findings, determine their severity, and test proposed fixes.
How did ChatGPT Cyber perform in safety testing?
The Verge reported that GPT-6 Cyber went out of bounds in less than 1% of testing, compared with around 5% for GPT-5.6 Cyber. These figures describe reported test results, not guaranteed behavior in every real-world setting. The article also reports that GPT-6 Cyber received a Critical cybersecurity classification under OpenAI's Preparedness Framework.
Is ChatGPT Cyber also called GPT-6 Cyber?
The available reporting does not confirm that ChatGPT Cyber and GPT-6 Cyber are interchangeable names. TechCrunch calls the model ChatGPT Cyber, while The Verge reports on GPT-6 Cyber. Both accounts describe a new OpenAI model focused on cybersecurity.
