OpenAI's New ChatGPT Cyber Looks for Software Flaws, but Only for Vetted Users

OpenAI announced ChatGPT Cyber on November 5, 2026. It is an AI model built for cybersecurity work, and the company says it will be available to its highest-tier users in the Daybreak Red program. A TechCrunch report on the announcement also cites a $1 billion initiative to support Daybreak cybersecurity professionals working on critical infrastructure.
A model is the AI system that does the work behind a product. In this case, the work is searching for software flaws.
According to the company, ChatGPT Cyber found 81 new vulnerabilities in Firefox, the Mozilla web browser. A vulnerability is a weak spot that an attacker could use. The model was also used to scan the Linux kernel, the core software of the Linux operating system. There it flagged 948 potential issues while staff prepared fixes.
These figures come from OpenAI and have not been checked by an independent party. Both codebases are large and have been around for years. Every finding still goes through human review and testing before it becomes a fix.
Daybreak is OpenAI's cybersecurity program. It combines the company's AI with security products, a network of partners, and support for defenders. The Daybreak page describes a system built on OpenAI models and Codex Security. It finds threats, writes patches, and checks that the fixes work, across both code and computer systems. Red is the top access tier.
Access is restricted. In an August 10 post, OpenAI said approved Daybreak partners can use its most advanced cyber models to deliver authorized, governed security services to customers. This is not a tool that anyone can sign up for. Partners are the channel that delivers the work, and they must be vetted first.
The models have a history of their own. OpenAI runs a program called Trusted Access for Cyber. In April 2026, the company expanded that program and introduced GPT-5.4-Cyber to vetted defenders. OpenAI calls GPT-5.6 Cyber its most advanced cybersecurity model for authorized vulnerability research and security testing.
Fortune reported on September 24 that GPT-6 Cyber was the fourth cybersecurity-focused model OpenAI released in 2026. The same day, Reuters, citing Fortune, reported that OpenAI was set to preview GPT-6 Cyber within days. ChatGPT Cyber is the name used in the November 5 announcement.
Astra is the next step, and it has not been released. OpenAI shared preliminary cybersecurity test results for Astra on August 7. A September 1 post said Astra is the first OpenAI model to meet the Critical cybersecurity capability threshold under the company's Preparedness Framework, its internal rulebook for judging risk. Under that framework, Critical means OpenAI's own process treats a model as capable of serious cyber harm. That is why the access limits matter.
The other side of the market is also moving. Japan's SoftBank Group launched a cybersecurity product based on OpenAI models in June 2026. The company said it is designed to counter breaches enabled by artificial intelligence, as Reuters reported. That means OpenAI's models are also running inside a security product sold by a separate company.
Questions have also been raised about OpenAI itself. On November 3, Reuters reported, citing sources, that OpenAI employees had warned for years before November 2026 about lapses in operational security. Those warnings concern how OpenAI protects its own systems. That is a separate matter from what its models can do, but the two meet in the Daybreak design. A tiered system is only as strong as the controls around its top tier.
The broader context here is a problem experts call dual use. The same skill that finds a flaw can also be used to exploit it. Think of a locksmith who can open any door for the owner and could just as easily open it for a burglar. The tiered model tries to give the skill to defenders and keep it from attackers. Whether vetting, partner oversight, and monitoring hold up at a larger scale is not something these announcements can settle. OpenAI has not yet shown that at the scale Astra would require.
In my view, the practical question is more ordinary. The Firefox and Linux figures suggest that AI can turn up more problems than people would find alone. That means more work for the people who sort through the results and ship the fixes, and maintainers who already work through long lists will feel that first. The upside is real. Defenders who fix more hidden bugs before attackers find them gain an advantage, and Daybreak is built around that goal. What is still open is whether the access controls keep pace as the models become more capable.


