# OpenAI rates its new GPT-6 Cyber model "Critical" for cyber risk and limits it to vetted security teams

OpenAI reports a 48.5% score on its ExploitGym benchmark against 32.1% for GPT-5.6 Cyber, along with working exploits for 948 Linux kernel bugs, and no outside lab has reproduced any of these figures.

- Published: 2026-11-06T10:11:19.661Z
- Canonical: https://polylog.news/ai/2026-11-06/openai-rates-its-new-gpt-6-cyber-model-critical-for-cyber-ri
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [OpenAI News](https://openai.com/index/introducing-gpt-6-cyber), [Polylog editors](https://polylog.news), [Fortune](https://fortune.com/2026/11/05/openai-debuts-new-cybersecurity-focused-ai-model-amid-growing-alarm-about-ai-driven-cyber-attacks/), [BankInfoSecurity](https://www.bankinfosecurity.com/openai-unveils-gpt-6-cyber-stronger-exploit-skills-a-33314), [CyberScoop](https://cyberscoop.com/openai-gpt-6-cyber-security-model/), [The Next Web](https://thenextweb.com/news/openai-gpt-6-cyber-launch-firefox-linux-daybreak-red)

OpenAI released [GPT-6 Cyber](https://openai.com/index/introducing-gpt-6-cyber) on November 5. The company describes it as a model for finding exploitable vulnerabilities and writing fixes, trained to stay inside the scope an operator authorizes. A Russian-language AI channel that tracks lab releases [summarized it more directly](https://t.me/ai_machinelearning_big_data/11295): the model searches for unknown vulnerabilities, analyzes code, and develops exploits on its own.

The headline number comes from OpenAI itself. GPT-6 Cyber scores [48.5% on ExploitGym](https://www.bankinfosecurity.com/openai-unveils-gpt-6-cyber-stronger-exploit-skills-a-33314), a test of whether a model can exploit a known flaw on a remote target. GPT-5.6 Cyber, released in August, scored 32.1%. OpenAI also says the model beats its general-purpose flagship, GPT-6 Astra, on most cyber tasks. Coverage disagrees on the Astra baseline. Some outlets cite a figure that appears to come from a separate web penetration-testing benchmark, so the Astra gap is not settled.

The claims from field testing go further. According to [Fortune](https://fortune.com/2026/11/05/openai-debuts-new-cybersecurity-focused-ai-model-amid-growing-alarm-about-ai-driven-cyber-attacks/), researchers used the model to find 81 high-severity Firefox flaws and to build working exploits for 948 Linux kernel vulnerabilities, 588 of them previously unknown. Reports differ on whether the 588 are a subset of the 948. Neither Mozilla's advisories nor the kernel maintainers' patch logs, as cited in this coverage, yet show the full count.

OpenAI moved the model's cybersecurity rating [from "High" to "Critical"](https://thenextweb.com/news/openai-gpt-6-cyber-launch-firefox-linux-daybreak-red) under its Preparedness Framework. Access goes first to [a small group of approved teams in the Daybreak Red program](https://cyberscoop.com/openai-gpt-6-cyber-security-model/). OpenAI says it trained the model to [refuse less often than GPT-6 Astra](https://www.bankinfosecurity.com/openai-unveils-gpt-6-cyber-stronger-exploit-skills-a-33314) and to resist reward hacking, meaning the model should not report success on a task it did not actually complete.

The release rests on two claims that together raise the commercial value of access: that the model is dangerous enough to require OpenAI's highest risk rating, and that it is safe enough to give to selected customers with fewer refusals. The confirmed facts so far are the release itself, the rating, and the restricted access. The benchmark results and the exploit counts remain unverified vendor assertions.

## What this means

Offensive exploit capability is now gated behind one vendor's access list, which determines who can use it and who cannot. The teams OpenAI admits get faster vulnerability discovery. Open-source maintainers such as Mozilla and the Linux kernel community get a surge of reports they did not choose to receive. Security vendors gain a stronger case that demand for their products will grow. That dynamic has already moved stock prices once: on September 14, after frontier-lab CEOs issued safety warnings, [CrowdStrike and Zscaler rose 12% and Palo Alto Networks 11%](https://247wallst.com/investing/2026/09/14/cybersecurity-stocks-surge-as-ai-safety-warnings-spark-security-bid-crowdstrike-and-zscaler-jump-12-palo-alto-rallies-11/). The consensus treats the gate as the safety control. The weaker point is how long the gate lasts, because open-weight models now score within a few points of closed models on agentic coding tests (see today's Kimi K3.1 Pro report). The likelier outcome is a measurable rise in kernel and browser security advisories credited to AI-assisted discovery within weeks, and a rival lab releasing its own gated cyber model within months. Both depend on OpenAI's counts holding up when maintainers publish. If those advisories show far fewer distinct or high-severity bugs than OpenAI claims, the read is wrong.

## What to watch

- Mozilla security advisories and Linux kernel stable patch notes over the next several weeks. A jump in fixes credited to AI-assisted discovery would confirm OpenAI's counts, and a small number would mark them as inflated.
- Whether Anthropic or Google releases a comparable cyber-specialized model with similar gated access. Matching releases would mean labs now compete on offensive capability for paying defenders.
- Whether any government asks to control who receives Daybreak Red access. That would show states treating model access as a controlled strategic asset.