Key takeaways
- GPT-6 Cyber surfaced 81 previously unknown Firefox vulnerabilities that Mozilla rated high severity, and Mozilla has already fixed many of them.
- On OpenAI's ExploitGym benchmark the model scored 48.5%, up from 32.1% for GPT-5.6 Cyber, though these are company figures rather than independent tests.
- Access starts with approved Daybreak Red customers through the API and Codex, with identity checks, misuse monitoring and human review for risky workflows.
In this story 4 sections
OpenAI's new GPT-6 Cyber model found 81 previously unknown security flaws in Firefox, and the company says it is the first model it rates "Critical" for cybersecurity under its own Preparedness Framework. OpenAI announced the model on Thursday, November 5, in a post on its site, and for now only a short list of vetted security teams can use it.
GPT-6 Cyber is a version of GPT-6 Astra trained for one job: finding software vulnerabilities, testing how an attacker could exploit them, and helping write the fixes. According to Help Net Security, OpenAI says the new model beat GPT-6 Astra on most cybersecurity tasks. CyberScoop quotes the company describing it as "a major step up from GPT-5.6 Cyber."
81 Firefox bugs, 948 Linux exploits and a Brotli flaw
The headline result comes from Mozilla's browser. GPT-6 Cyber surfaced 81 new vulnerabilities in Firefox that Mozilla classified as high severity, Help Net Security reported. Mozilla has already shipped fixes for many of them, and the model helped by proposing patches.
The second result is about Linux. OpenAI says the system developed working exploits for 948 Linux kernel vulnerabilities, a finding the company detailed in a separate writeup the day before launch. It also helped validate a critical remote code execution bug in Brotli, the compression library used widely across the web. CyberScoop reported that all of the flaws were disclosed to their owners and patched.
On OpenAI's own tests, the jump over the previous specialist model is large:
- ExploitGym: 48.5% success for GPT-6 Cyber, compared with 32.1% for GPT-5.6 Cyber.
- Honeypot test: the model stayed away from an out of scope service in 99.93% of runs with no safety instructions, and in 100% of runs when those instructions were included.
Those are company benchmarks, not independent audits, so treat them as OpenAI's claims until outside researchers publish their own numbers.
Daybreak Red customers get GPT-6 Cyber first
You cannot sign up for GPT-6 Cyber today. The model is going first to a limited group of approved customers in OpenAI's Daybreak Red program, through the API and Codex. CyberScoop describes Daybreak Red as the tier that uses more advanced or nonpublic security models for offensive red teaming, meaning attacking systems on purpose to find weak spots before criminals do. OpenAI says it plans to open the model to everyone in that program later. Neither outlet reported a price or a date for wider release.
A separate track, Daybreak Blue, sits inside OpenAI's Trusted Access Program for defenders. That tier ran into its own trouble this week. CyberScoop reported that OpenAI told members their Daybreak Blue sessions on personal accounts would be used to train models even if they had opted out. After complaints, OpenAI safety lead Fouad Matin said customers could keep opting out of training.
The "Critical" label and the guardrails around it
A Critical rating in OpenAI's framework signals the company believes the model could meaningfully help someone carry out serious cyberattacks if it were misused. That is why the release comes wrapped in restrictions rather than a public chat window.
According to Help Net Security and CyberScoop, the safeguards include:
- Monitoring for actions that go beyond what a user is authorized to test.
- Extra identity and account security checks for people with access.
- Protection against stolen credentials being used to reach the model.
- Checks that block high risk cyber help, plus sandboxing and human review for the riskiest workflows.
- More frequent check ins from the model when someone asks for high risk behavior.
OpenAI also named industrial security firm Dragos as a partner. Dragos tested the protections in a controlled lab using Daybreak Red models and, according to Help Net Security, found two zero days in engineering software during that work.
How GPT-6 Cyber fits a busy month for AI security
The launch lands in a week when security vendors and AI labs are both racing to put guardrails on agents. CrowdStrike recently added a way for admins to shut down rogue AI agents, and Anthropic opened a tool that proves which model wrote a reply. Regulators are moving too, with China drafting rules that force AI firms to disclose safety incidents.
The bigger question is what happens when models this good at breaking software become common. Labs have started to say so plainly; Anthropic's IPO filing included a warning about AI systems improving themselves. For now, OpenAI is betting that keeping GPT-6 Cyber behind vetted access lets defenders patch faster than attackers can catch up. Whether that gap holds will depend on how quickly rival labs ship similar tools without the same gates.
For more coverage of models, policy and safety, see our AI news section.
Share this story
Spotted an error? Tell us
