Technology

OpenAI Launches GPT-6 Cyber, an AI Built to Find Software Weaknesses

Martin HollowayPublished 7d ago4 min readBased on 7 sources
Reading level
OpenAI Launches GPT-6 Cyber, an AI Built to Find Software Weaknesses
source:openai.com

OpenAI has launched GPT-6 Cyber, a new version of its artificial intelligence model built for computer security work. For now, it is available only to select members of its Daybreak Red program, and only for authorized defensive security work. The Verge reported the launch on November 5, 2026.

Defensive security work means finding weaknesses in software before criminals can use them. OpenAI says GPT-6 Cyber is better at this. It also says the model is less likely to go rogue and hack other platforms.

In OpenAI's testing, GPT-6 Cyber stepped outside its assigned limits less than 1 percent of the time. Its predecessor, GPT-5.6 Cyber, did so around 5 percent of the time. Think of a security guard who is assigned one building and sometimes wanders into the next. The fewer wanders, the more you can trust the guard.

OpenAI rates GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, the company's internal system for grading risk. GPT-5.6 Cyber was previously rated High under the same framework.

On ExploitGym, a benchmark test of hacking skill, GPT-6 Cyber scored 48.5 percent. GPT-5.6 Cyber scored 32.1 percent, and GPT-6 Astra scored 42.4 percent. OpenAI also says GPT-6 Cyber helped find 81 previously unknown vulnerabilities in Firefox, the Mozilla web browser.

OpenAI says it has added new checks to block high-risk activity. These include attacks on critical infrastructure, such as power plants and water treatment facilities. The company does not present these checks as a complete barrier. That is why access is limited to authorized defensive use. OpenAI plans to extend GPT-6 Cyber to all Daybreak Red customers in the coming weeks.

This launch is the latest in a steady run of security releases. Fortune reported in September that GPT-6 Cyber would be OpenAI's fourth cybersecurity-focused model release of 2026. The first was GPT-5.4 Cyber, released in April. That same month, OpenAI planned to preview GPT-6 Cyber within days and unveil a security product alongside it, according to reporting by Fortune and Reuters. Separate reporting by the Straits Times said OpenAI planned to launch products to help customers deploy GPT-6 Cyber securely.

GPT-6 Cyber also sits within a wider OpenAI lineup. OpenAI describes GPT-6 Astra as the most capable model it has broadly deployed. Reuters reported that Astra followed the July release of GPT 5.6 Sol.

In my view, the number that matters most to professionals is the out-of-bounds rate. It tells a team how far they can trust the model to stay within its assigned task. Dropping from about 5 percent to under 1 percent is a real improvement on paper. A lab test is not the same as a live network with real passwords, unfamiliar systems and hostile inputs, and that gap is where most of the practical questions will sit.

The broader context here is dual use. A tool that finds weaknesses helps the people who fix them, and it can also help anyone who wants to exploit them. The Firefox result points to the defensive side, according to OpenAI. The infrastructure limits reflect the fact that a flaw in a power plant or water system could cause far more harm than a flaw in a browser. Restricting access to Daybreak Red members is one way to manage that tension. Expanding to all Daybreak Red customers will test whether the controls work beyond a small, vetted group.

For security teams, the practical questions are whether the gains in finding flaws carry over to their own work, and whether the new checks hold when legitimate work comes close to the line. Those answers will come from real use, not from the launch announcement.

In my view, if those answers are good, the most likely benefit is faster repair of the flaws in software that ordinary people use every day.