A specialised release, not general access

OpenAI introduced GPT-6 Cyber on November 5, a version of GPT-6 Astra trained for authorised security work. Access begins with selected approved Daybreak Red customers through the API and Codex, rather than an unrestricted public rollout.

The company says it intends to expand availability to all Daybreak Red customers. That is a rollout plan, not a statement that every applicant or ordinary subscriber can use the model now.

Capability is not one score

OpenAI reports a 48.5% result on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. This is evidence from a defined company evaluation, not a measure of how many vulnerabilities an arbitrary organisation will find or fix.

The system card also reports external evaluator Irregular’s results. GPT-6 Cyber completed all 11 CyScenarioBench challenges at least once, versus eight for Astra. On FrontierCyber, however, it solved 90 of 288 challenges, below Astra’s 96. Neither model solved the 15 hardest challenges, and the evaluator observed no successful attacks on fully hardened targets.

These comparisons describe particular tasks and test conditions. They do not support a claim that the specialised model is better at every security job.

Evidence of fixes has a different scope

Mozilla’s October 27 Firefox 159 advisory credits OpenAI researcher Amy Burnett for multiple reported flaws. That is a primary record of credited vulnerability reports and fixes. It does not independently establish GPT-6 Cyber’s contribution to each finding or validate OpenAI’s overall benchmark claims.

The boundary still matters

The system card describes identity and account-security requirements, monitoring and default API network allowlisting. It recommends isolated environments with explicit limits on targets, credentials, network access and permitted actions, plus human oversight for high-risk work.

For a security team, the useful question is not simply whether an agent can finish a test. It is whether its actions stayed within written permission and whether a person can inspect the evidence and stop the work. A reported capability gain does not remove those separate decisions.

Sources

  1. OpenAI: Introducing GPT-6 CyberNovember 5 launch announcement, selected-customer rollout and company-reported ExploitGym comparison. Not an independent field trial.
  2. OpenAI: GPT-6 Cyber System CardNovember 5 first-party safety report, including attributed Irregular evaluations, updated benchmark comparisons and access/oversight safeguards. No operational exploitation guidance reproduced.
  3. Mozilla: Security vulnerabilities fixed in Firefox 159October 27 primary advisory and reporter credits, including Amy Burnett of OpenAI. Does not independently attribute findings to a particular model or corroborate aggregate model-performance claims.