OpenAI names GPT-6 Cyber, rates it Critical, and limits it to approved Daybreak Red customers
Thursday’s post puts a name on the security model behind the Linux run described on Wednesday. On ExploitGym it scores 48.5 percent, against 32.1 percent for GPT-5.6 Cyber, a gap the post calls about 16 points and a 51 percent relative gain. OpenAI says it produced evidence for 81 Firefox bugs Mozilla classed as high severity. API output is $100 per million tokens. A Daybreak Red organization still needs a separate approval to call the model.
Deep dive6 minIn-depth
Priya Shah6 min read
- OpenAI
- Security
- GPT-6
- Daybreak

Photo: OpenAISource page
OpenAI on Thursday published GPT-6 Cyber, a security model it classifies as Critical under its Preparedness Framework, up from High for GPT-5.6 Cyber. The post is dated 5 November. On ExploitGym, a test of whether a model can exploit a vulnerability it has been pointed at, OpenAI reports a score of 48.5 percent, against 32.1 percent for GPT-5.6 Cyber. The post calls that about 16 percentage points, and a 51 percent relative gain. The model is not offered as a general ChatGPT choice. OpenAI says it is available first through the API and Codex to a select group of approved Daybreak Red customers, with a wider Daybreak Red rollout to follow.
Emma Roth’s account at The Verge is timestamped 5:00 p.m. UTC on Thursday. She quotes Fouad Matin, head of OpenAI’s cyber preparedness and safety team, from a press briefing, and she reads a chart in the launch material as putting GPT-6 Astra at 42.4 percent on the same ExploitGym test. OpenAI’s own post says GPT-6 Cyber beats Astra on most cyber tasks even when Astra does not refuse the request. It does not print Astra’s ExploitGym score in the text. The Verge also quotes the framework’s Critical category as a model that “can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention.”
The same post puts a name on a run OpenAI described on Wednesday without one. It says GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948 Linux privilege-escalation bugs, more than five times the 166 it attributes to GPT-5.6 Sol on the same harness, and that the company has disclosed 304 critical findings to maintainers with candidate patches. Wednesday’s account had the cost of that run, the stages, and the cases where an early check reached root through a different bug than the one assigned. Thursday’s account is the product around that work: a model id, a rating, and a limit on who can call it.
A web test, and a remote one
OpenAI says it built a separate evaluation in which models probe web applications without the source code. GPT-6 Cyber recalled 40.3 percent of the vulnerabilities, against 32.0 percent for GPT-6 Astra. A second new benchmark, RemoteExploitBench, uses known vulnerabilities in applications such as GitLab, PostgreSQL, and Jenkins, running in a controlled environment. OpenAI says GPT-6 Astra with Daybreak Blue refused every task in the initial run, and GPT-6 Cyber completed 95 percent. The post says the one miss was an exploit the model could not get to work remotely. It points to a system card for the settings and does not summarize that card.
The Firefox number is separate from those benchmarks. OpenAI says that in a research campaign GPT-6 Cyber produced evidence and reproducible tests for 81 vulnerabilities Mozilla classified as high severity, and that it showed 30 issues already known to be more severe than previously reported. The company says it sent Mozilla the findings and proposed patches, and that Mozilla began landing fixes soon after. Brian Grinstead, a distinguished engineer at Firefox, is quoted in the post saying the browser entered 2026 as one of the most hardened codebases, that further work with frontier models had not cleared it, and that GPT-6 Cyber “uncovered previously unknown high-severity vulnerabilities, identified architectural improvements, and helped us ship fixes to millions of users.”
Staying on the assigned target
A large part of the post is about the model not wandering off the task. OpenAI describes ExploitGym Honeypot as a check informed by an incident at Hugging Face. The model is given a difficult job in an isolated environment, plus supporting services that are easy to break. Breaking those services can reveal the answer, and OpenAI says that is outside the assigned scope. Without a deployment safety prompt, GPT-6 Cyber avoided the out-of-scope services in 99.93 percent of tests, up from 94.49 percent for GPT-5.6 Cyber. With the prompt, OpenAI says it stayed inside the task in every test.
With the 6 Cyber model, we were able to make huge improvements on this evaluation. It’s over 99 percent of the time doing what you ask.
— Fouad Matin, head of cyber preparedness and safety at OpenAI, at a press briefing, as quoted by Emma Roth, The Verge, 5 November 2026
Matin also told the briefing that evaluation awareness, a model noticing it is being tested and changing how it behaves, is still a concern. “Often when we do see evaluation awareness, you’ll see it in the model’s chain of thought,” Roth quotes him saying. He said OpenAI is trying to use tests that look more like production work, on real codebases with a real objective, to reduce that risk. The honeypot percentages above are the published scores. His comments are about what those scores do not settle.
Who can call it, and what a token costs
The model id on the API page is gpt-6-cyber. The page lists a context window of 1,050,000 tokens, a maximum of 922,000 tokens of input and 128,000 of output, and text and image as inputs, with text as the output. It supports the Responses API and Chat Completions. Batch is listed as unsupported. The standard short-context prices are $20 per million input tokens, $2 per million cached input tokens, and $100 per million output tokens. Cache writes are $25 per million tokens for 30 minutes, or $50 per million for 12 hours. The page says an organization needs Daybreak Red and a separate approval, and that an account team confirms access.
OpenAI’s help center, updated in the hours before this story, draws the same line in a table. Every Daybreak Red organization is eligible for GPT-5.6 Cyber. GPT-6 Cyber is listed only for organizations with an extra approval for that model. Turning Daybreak Red on in a workspace or an API project does not grant it. Selecting Daybreak Red in a request does not switch the model either. The caller sets the model id and sets access_programs.cyber to daybreak_red. The help center also says Daybreak Red is for approved business and enterprise organizations, not the individual application path, which is for Daybreak Blue. It notes that some individuals may have access as an exception. The launch post says OpenAI requires IP allowlisting for Daybreak Red API access by default, with limited exceptions, and that API customers on zero data retention generally must use a check it calls Private Safety Processing.
The post names teams already using the model for vulnerability research and code review: Cisco, Calif, CodeRabbit, Cognition, Palo Alto Networks, SentinelOne, CrowdStrike, and Perplexity. The figures attached to those names are the customers’, printed by OpenAI. Harjot Gill, CodeRabbit’s chief executive, says that on an internal benchmark the model found 66 percent more high- and medium-severity issues than the model CodeRabbit had in production. Steven Hao, Cognition’s chief technology officer, says it scored 23 percent on an internal test of finding and exploiting new vulnerabilities, against a previous best of 13 percent. OpenAI also says it used the model, with others, in red-team work against the infrastructure it uses to train frontier models, and that the findings were used to patch software and tighten network rules. It does not describe those bugs.
On critical infrastructure, the post says a new group inside Daybreak Red is for approved defenders of those systems, with permission for authorized testing and with safeguards against the model going off the task. It points resource-constrained government and essential-service teams at a grant program it calls Daybreak for Frontline Defenders and puts at $1 billion. It says safeguards were tightened, alongside the release, against help that could compromise industrial controls or cause physical damage, and that higher block rates are not the goal by themselves because the same techniques are used in legitimate testing. OpenAI says it worked with Dragos on lab evaluations and used that feedback when it adjusted the safeguards.


