OpenAI launches GPT-6 Cyber for vetted Daybreak Red teams, rated Critical for cyber capability

What OpenAI released

OpenAI introduced GPT-6 Cyber on 5 November 2026, describing it as a model for finding exploitable vulnerabilities and developing fixes that is trained to stay within the scope a customer has authorized. GPT-6 Cyber is the successor to GPT-5.6 Cyber and sits in the same GPT-6 generation as GPT-6 Astra, the general model OpenAI released in September.

Access is narrow. OpenAI says the model is available first to a select group of approved customers in Daybreak Red, the more permissive of the two vetted tiers in its Daybreak trusted-access programme, through the API and Codex. The company says it will widen access to all Red customers over the coming weeks. Daybreak Red covers offensive-style work such as simulated attacks and red teaming, while the separate Daybreak Blue tier is meant for defensive work. Coverage by CyberScoop, Fortune and The Next Web reports the same access terms. OpenAI has not published a list price or a date for general availability; figures that circulated after the launch do not match across outlets and are left out here.

The Firefox and Linux results OpenAI cites

OpenAI's headline claim is that GPT-6 Cyber found 81 previously unknown high-severity vulnerabilities in Mozilla Firefox and helped write candidate patches, and that Mozilla has already shipped fixes for many of them. The announcement does not say how those 81 overlap with the fixes Mozilla released in its regular October update, so the share of the bugs still open is not clear from the post.

The second result is the Linux kernel run OpenAI described a day earlier. In that run an autonomous pipeline reproduced 1,724 kernel vulnerabilities and built working local privilege escalation exploits for 948 of them, at a cost of about $1.2 million in API credits. Our report on that run noted that OpenAI had not named the model; the launch post now ties those numbers to GPT-6 Cyber. OpenAI says fixes are going to kernel maintainers in order of severity, with the rest routed through its Patch the Planet partners.

On benchmarks, OpenAI reports 48.5% on ExploitGym, which tests whether a model can exploit a known flaw, against 32.1% for GPT-5.6 Cyber. These are vendor figures. No independent lab had published its own evaluation of the model by the time of writing.

Why OpenAI rates it Critical and how it limits use

OpenAI classifies GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, the highest tier, compared with High for GPT-5.6 Cyber. The rating describes what the model can do, not an incident. GPT-6 Astra was the first OpenAI model to reach that level, as our September report covered; a model tuned for exploitation work at the same tier explains why access sits behind vetting rather than a public API.

The main safety claim is about scope. OpenAI says that in tests without a deployment safety prompt, the model avoided out-of-scope services placed next to its target 99.93% of the time, compared with 94.49% for GPT-5.6 Cyber, and that with the prompt in place it stayed in bounds in every test. The tests use a honeypot environment with easy-to-compromise services that the model is not authorized to touch. Even a small failure rate matters at scale: 0.07% of a very large number of automated actions is still a real number of actions against systems nobody approved.

Reported controls around the deployment include authentication of Daybreak customers, IP allowlisting, sandboxed execution and monitoring of model behaviour, alongside restrictions on work against critical infrastructure. The approach mirrors how OpenAI handled GPT-5.6 Cyber, but the higher rating means the vetting carries more weight: the same capabilities that let a red team prove a bug is exploitable would help an attacker who got through it.

What it means for defenders and what to watch

For security teams, the practical change is that a frontier lab now sells exploit development as a supported, vetted product rather than a research demo. Vendors and open-source maintainers should expect more AI-generated vulnerability reports, many with working proof-of-concept code, arriving through programmes like Patch the Planet. The bottleneck shifts from finding bugs to triaging and fixing them, and the Firefox and Linux numbers suggest that queue will grow faster than many projects can staff for. Anthropic has made similar claims about Claude building exploit chains, and Google's threat intelligence group has reported AI-assisted discovery pushing up disclosure volumes, so the trend is not limited to one company; our AI cybersecurity section tracks these reports.

Open questions remain. OpenAI has not said how many of the Firefox and Linux findings have been fixed and assigned CVE numbers, what GPT-6 Cyber costs, or when Daybreak Red customers outside the first group will get access. The scope figures come from OpenAI's own honeypot tests; an outside evaluation, for example from the UK AI Security Institute, would show whether they hold in other environments. The next signals to watch are Mozilla's and the kernel maintainers' advisories crediting the model, and whether OpenAI extends any version of it to Daybreak Blue defenders.

Questions people ask

Who can use GPT-6 Cyber?

Only approved customers in OpenAI's Daybreak Red tier, starting with a select group and widening to all Red customers over the coming weeks, according to OpenAI. It is not in ChatGPT or the public API.

What does the Critical rating mean?

It is the highest cybersecurity capability level in OpenAI's Preparedness Framework. It describes what the model can do and is the reason OpenAI keeps access behind vetting.

How much does GPT-6 Cyber cost?

OpenAI has not published a price in its announcement. Figures reported by some outlets after the launch do not match each other.

Sources

1. GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents (OpenAI, 5 Nov 2026)
2. OpenAI debuts GPT-6 Cyber, touting better performance and safety guardrails (CyberScoop, 5 Nov 2026)
3. OpenAI debuts new cybersecurity-focused AI model amid growing alarm about AI-driven cyber attacks (Fortune, 5 Nov 2026)
4. OpenAI's GPT-6 Cyber found 81 new Firefox flaws and is rated Critical (The Next Web, 5 Nov 2026)
5. OpenAI's GPT-6 Cyber ups the ante on vulnerability discovery (CSO Online, 5 Nov 2026)