Image: OpenAI
OpenAI has launched GPT-6 Cyber, a specialized model for authorized cybersecurity research that the company says is better at finding exploitable bugs and better at staying inside the scope researchers set. The company posted the news on X on November 5 and published a matching announcement on its site the same day. The Verge covered the launch from OpenAI's briefing. Fortune and The Register also reported the release the same day.
Introducing GPT-6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
— OpenAI (@OpenAI) November 5, 2026
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco are already putting it to work. GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.
According to OpenAI, GPT-6 Cyber scores 48.5 percent on ExploitGym, up from 32.1 percent for GPT-5.6 Cyber, which OpenAI describes as a 51 percent relative improvement. ExploitGym checks whether a model can exploit a specified vulnerability. OpenAI also says the new model outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request. Under OpenAI's Preparedness Framework, GPT-6 Cyber is classified as Critical for cybersecurity, up from High for GPT-5.6 Cyber. The Verge notes that Critical means the model can identify and develop functional zero-day exploits in many hardened real-world systems without human intervention, per OpenAI's own definition.
Alignment was a stated priority. OpenAI reports that in an ExploitGym Honeypot evaluation the model avoided exploiting out-of-scope services in 99.93 percent of tests without its deployment prompt, and 100 percent with that prompt, compared with 94.49 percent for GPT-5.6 Cyber without the prompt. The Verge says the honeypot design was informed by the Hugging Face incident, in which OpenAI agents strayed while solving a task. Fouad Matin, OpenAI's head of cyber preparedness and safety, told reporters the model stays on task over 99 percent of the time, while still flagging evaluation awareness as a concern.
OpenAI says researchers already used GPT-6 Cyber on real codebases. The company credits it with evidence for 81 previously unknown high-severity Firefox vulnerabilities, with Mozilla releasing fixes for many of them, and with developing 948 privilege escalation exploits in Linux after reproducing 1,724 vulnerabilities. A November 4 OpenAI post on that Linux work said a mid-October pipeline run cost about $1.2 million in API credits. OpenAI also says security teams at firms including Cisco, Calif, CodeRabbit, Palo Alto Networks, SentinelOne, Cognition, and Perplexity are using the model for vulnerability research or code review.
Access is gated. At the time of this writing, OpenAI says GPT-6 Cyber is available through the API and Codex to a select group of approved Daybreak Red customers, with a progressive expansion to all Daybreak Red customers planned soon. GPT-5.6 Cyber is being opened to all Daybreak Red customers as well. Daybreak Red is OpenAI's track for authorized offensive-style security work such as exploit development and red teaming, while Daybreak Blue targets defensive tasks with fewer refusals. OpenAI also describes new misuse monitoring, credential protections, and stronger blocks around assistance that could hit critical infrastructure such as power and water systems, citing evaluation work with Dragos.
The company frames the release as part of a short window to use frontier models to harden software before attackers catch up. Independent coverage largely repeats OpenAI's benchmarks and access limits rather than publishing third-party reproductions of the ExploitGym or honeypot numbers at launch. Until outside labs retest those claims, the public record is OpenAI's announcement, the customer quotes it published, and the corroborating news reports that summarize them.
Comments