OpenAI has released GPT‑6 Cyber, a restricted cybersecurity model that the company says has already uncovered serious vulnerabilities in widely used software, including 81 previously unknown high-severity flaws in Firefox and working exploits for 948 Linux kernel local privilege escalation vulnerabilities.
A restricted model with offensive capability
GPT‑6 Cyber is a specialized version of GPT‑6 Astra, trained for authorized vulnerability discovery, exploit development, web penetration testing and fix verification. In its system card, OpenAI gives the model a Critical cybersecurity capability designation under its Preparedness Framework, while saying access requires identity verification, use restrictions and organization-level security controls.
The model is designed to reduce unnecessary refusals for approved security work while still blocking harmful requests. That distinction is central to the release because the same skills needed to verify a vulnerability can also make exploitation easier if access controls and safeguards fail.
What OpenAI says it found
The most eye-catching claim is Firefox. OpenAI says GPT‑6 Cyber produced evidence and reproducible tests for 81 novel Firefox vulnerabilities classified by Mozilla as high severity, and showed that 30 known issues were more severe than previously reported. According to OpenAI, Mozilla has released fixes for many of those Firefox findings.
Mozilla’s Firefox 159 security advisory, announced Oct. 27, lists high-impact fixes across areas such as sandbox escapes, privilege escalation, use-after-free bugs, information disclosure and boundary-condition flaws. Not every entry in that advisory is attributed to OpenAI; Mozilla lists different reporters across the disclosed CVEs.
The Linux work came from a separate OpenAI research post published Nov. 4. That post described the system as powered by an upcoming cyber model; the following GPT‑6 Cyber launch named GPT‑6 Cyber as the model behind the 948-exploit work. In the Linux campaign, OpenAI says the pipeline identified 2,308 candidate findings, dynamically reproduced 1,724 vulnerabilities and produced 948 working local privilege escalation exploits on a commonly used Linux distribution.
OpenAI said the Linux run cost US$1,213,430 in API credits, or US$1,279 per local privilege escalation exploit. The company also said it followed guidance from the Linux kernel community by starting with critical vulnerabilities, disclosing reports with candidate patches and continuing work with partners to report and fix remaining issues.
The benchmark numbers are strong, but bounded
OpenAI says GPT‑6 Cyber scored 48.5% on ExploitGym, compared with 32.1% for GPT‑5.6 Cyber. In a web penetration-testing evaluation, it reported 40.3% recall, compared with 32.0% for GPT‑6 Astra. On RemoteExploitBench, a controlled benchmark using known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins, OpenAI says GPT‑6 Cyber completed 95% of tasks.
Those numbers do not measure ordinary internet attack success. They measure controlled tasks designed to test vulnerability discovery, exploit development and scope-following. OpenAI’s system card also warns that scores from older system cards should not always be treated as directly comparable because evaluation methods can change.
OpenAI says scope-following improved
OpenAI’s system card says GPT‑6 Cyber avoided exploiting out-of-scope services in 99.93% of ExploitGym Honeypot tests without deployment safety instructions and 100% with those instructions.
The system card still reports limits. Compared with GPT‑6 Astra, GPT‑6 Cyber stopped less often on warning-heavy tasks where stopping was appropriate. It was also flagged in 20.29% of blocker-reporting responses, mostly because its final answer did not repeat a qualification already disclosed earlier in the interaction.
Access is limited for now
The Nov. 5 launch positions GPT‑6 Cyber as a tool for vetted defenders, not a model available to ordinary ChatGPT users. OpenAI says initial access is limited to selected approved Daybreak Red customers through the API and Codex, with broader Daybreak Red access planned.
OpenAI’s API catalog lists the model ID as gpt-6-cyber, with a 1,050,000-token context window and 128,000 maximum output tokens. Standard short-context prices are US$20 per million input tokens and US$100 per million output tokens.
The pressure shifts from discovery to patching
The findings do not prove that every system is compromised. They do show that mature codebases can still contain enough latent weakness for a specialized model to generate a large review and patch queue.
OpenAI’s Linux write-up says vulnerability discovery is becoming easier to scale, while prioritization and validation remain operational bottlenecks. That matches the tension around GPT‑6 Cyber: finding candidate flaws is only useful if maintainers can confirm them, prioritize the dangerous ones and ship fixes without overwhelming review teams.
For software teams, the practical consequence is faster vulnerability cycles. Security review, dependency updates, patch validation and incident response are likely to need model-assisted rescan cycles, where newer systems revisit code that older systems already checked.
For everyday users, the immediate action is familiar: install browser and operating-system updates promptly. OpenAI’s findings do not show that all unpatched systems have been exploited, but they add one more reason not to delay fixes once vendors release them.
Get new small business insights by email
Practical ideas and useful articles to help you make better business decisions.
Tech Help Canada Staff researches, writes, and reviews practical content for business owners and professionals. Our coverage spans business, marketing, SEO, technology, and the tools and systems people use to grow and operate online. We focus on clear, useful information backed by research, hands-on experience, and editorial review. Learn more about our team and editorial standards. Need help with something? Contact Us
