TL;DR
On 5 November OpenAI released GPT-6 Cyber, a model built to find, exploit and help patch software vulnerabilities. OpenAI classes it as Critical for cybersecurity under its Preparedness Framework, a step up from the High rating given to GPT-5.6 Cyber. It is available only to a selected group of approved customers on Daybreak Red, OpenAI’s platform for authorised offensive security work.
What the model does
OpenAI’s headline number is ExploitGym, a test of whether a model can exploit a specified flaw in a remote target. GPT-6 Cyber scored 48.5%, against 32.1% for its predecessor. On a new benchmark using known bugs in tools such as GitLab and Jenkins, it completed 95% of tasks. GPT-6 Astra, OpenAI’s mainline model, refused every one of them when run under Daybreak Blue settings.
GPT-6 Cyber is also the unnamed model behind the 948 Linux privilege escalation exploits OpenAI described on 4 November. OpenAI says it reproduced 1,724 kernel vulnerabilities and disclosed 304 critical findings to maintainers. Separately, Mozilla classed 81 new Firefox flaws it found as high severity, and has begun releasing fixes.
Staying inside the lines
OpenAI puts as much weight on behaviour as on capability. In a honeypot test, the task environment includes deliberately easy targets the model is not authorised to touch. Without its deployment prompt, GPT-6 Cyber left them alone in 99.93% of runs, up from 94.49% for the previous model. With the prompt, it stayed in bounds every time. OpenAI says the test was informed by an earlier incident involving Hugging Face.
New safeguards include IP allowlisting for Daybreak Red API keys by default, monitoring of whether actions match a user’s authorisation, and blocks on help that could compromise industrial control systems. OT security firm Dragos tested OpenAI’s permissive Daybreak Red models on a simulated power plant. Its chief AI officer, Rob T. Lee, said a frontier model found two zero-days in protection-relay software with almost no guidance.
Why it matters for UK organisations
The launch follows the NCSC’s warning on AI-enabled attacks on UK infrastructure, and here a vendor has put a Critical rating on its own product. In our view, the Dragos result is the detail UK operators of energy and water systems should note: Lee says the skill needed for reconnaissance on those networks is falling. OpenAI is offering subsidised access through a $1 billion grant programme for resource-constrained defenders, including government bodies.
Looking Forward
OpenAI says every Daybreak Red customer will get access before long. All benchmark and alignment figures are its own, and the system card is the place to scrutinise them.