THE ESSENTIALS IN 30 SECONDS

  • GPT-6 Cyber is available through the API and Codex to a limited set of approved Daybreak Red customers, with wider Daybreak Red access planned over the coming weeks.
  • OpenAI credits the model with 81 previously unknown high-severity Firefox vulnerabilities and says it powered the Linux kernel exploit run published a day earlier.
  • The company rates the model Critical for cybersecurity; on its reported benchmarks it beats GPT-6 Astra on most cyber tasks but not all.

OpenAI introduced GPT-6 Cyber on November 5, a version of its GPT-6 Astra model specialised for authorised vulnerability discovery and exploit development. According to CyberScoop and CSO Online, the model is first going to a select group of approved customers in Daybreak Red, the more permissive of the two access tiers in OpenAI's trusted-access cyber programme, through the API and Codex. OpenAI says it will extend access to all Daybreak Red customers over the coming weeks. No pricing was reported.

What OpenAI says the model can do

OpenAI credits GPT-6 Cyber with finding 81 previously unknown high-severity vulnerabilities in Firefox; CSO Online reports that Mozilla has begun shipping fixes. The company also says the model powered the Linux kernel research it published on November 4, in which an agent pipeline produced exploits for 948 privilege-escalation flaws, and that it validated a critical remote code execution bug in the Brotli compression library. CyberScoop reports that the disclosed vulnerabilities were passed to their owners and patched.

On benchmarks cited by Unite.AI from OpenAI's materials, GPT-6 Cyber scored 48.5% on ExploitGym, against 42.4% for GPT-6 Astra and 32.1% for GPT-5.6 Cyber. The picture is not uniform: on Irregular's FrontierCyber set it solved 90 of 288 challenges, slightly fewer than Astra's 96. OpenAI rates the model Critical for cybersecurity under its Preparedness Framework, the same level as Astra.

Access controls and scope

Daybreak Red applicants go through identity and eligibility checks, and organisations must meet requirements such as single sign-on, multi-factor authentication and API IP allowlisting, Unite.AI reports. OpenAI says it adds more frequent check-ins for high-risk requests and extra monitoring, and can pause Codex runs or revoke access. A large part of the pitch is staying in scope: in a honeypot test, the model attacked an out-of-scope service in 0.07% of runs, versus 5.51% for GPT-5.6 Cyber. CSO Online quotes Noma Security's Chris Hughes noting that even a small rate could mean many unauthorised actions at scale.

Open questions

The performance figures come from OpenAI and its partners and have not been independently replicated. The launch also comes as the company faces criticism over a plan to use Daybreak Blue session data to improve its models; its safety lead told CyberScoop the data would serve abuse prevention rather than training on defenders' work.