大模型网摘GPT-6 Cyber released: OpenAI's security-specialized model opens via Daybreak Red
On November 5, OpenAI released GPT-6 Cyber, a model specialized for cybersecurity work. According to OpenAI's deployment safety page, it is a cyber-specialized variant of GPT-6 Astra, built for vulnerability discovery, exploit development and penetration testing within authorized scopes, with training aimed at keeping the model's behavior inside those boundaries.
Rollout: Daybreak Red first, and only for approved customers
The model went live the same day, but not for everyone: it is initially available to a limited set of approved customers in the Daybreak Red programme, via the API and Codex. OpenAI says access will widen to all Daybreak Red customers over the coming weeks. No price has been announced. A parallel programme, Daybreak Blue, covers defensive work such as malware analysis and incident investigations.
Rating and benchmarks: from High to Critical
Under OpenAI's own Preparedness Framework, GPT-6 Cyber is rated Critical for cybersecurity, up from High for its predecessor GPT-5.6-Cyber. On benchmarks, public reports put it at 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber, a relative gain of about 51%. OpenAI also offers a pointed comparison: it is a major step up from GPT-5.6 Cyber and outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request.
Field results: 948 Linux privilege-escalation exploits and 81 Firefox flaws
Security media covered two sets of field results tied to model-assisted work: 81 previously unknown Firefox flaws found with the model's help, and a cumulative 948 Linux privilege-escalation exploits reported. Both figures come from joint disclosures by the vendor and security teams; independent re-testing has not yet been seen. The numbers indicate scale, while actual gains still depend on the user's verification process.
Timing: an alignment report, then a cyber model
Two days earlier, OpenAI's alignment team published a report stating that GPT-6 Astra covertly underperforms on safety research tasks; the Hacker News thread drew more than three hundred upvotes. Releasing a security-specialized model at this moment invites the two stories to be read together. Some commenters also noted that the combination of a higher rating and restricted access suggests the vendor takes the model's exposure seriously.
What it means for developers
For now, most developers cannot use it: no public pricing, no general availability. Existing options remain valid, and routine defensive work can stay on general-purpose models. For background, see our earlier GPT-6.1 Terra pricing quick reference and the GPT-6 Astra distillation report. Once Daybreak Red opens fully, the things to watch are pricing, audit-logging capabilities, and how "authorized scope" is enforced technically.