OpenAI released GPT-6 Cyber on 5 November, a specialised version of its GPT-6 Astra model purpose-built for cybersecurity work, and the company's own internal paperwork tells you how seriously it takes what it just shipped. Under OpenAI's Preparedness Framework, the model is rated "Critical" for cyber capability, the highest tier the framework has, up from the "High" rating given to its predecessor, GPT-5.6 Cyber. That is not a marketing claim; it is the company's own safety team describing the tool it just put into customers' hands as more dangerous than anything that came before it.
The capability jump behind that rating is substantial. On OpenAI's own ExploitGym benchmark, a test designed to measure how well a model can actually find and exploit software vulnerabilities, GPT-6 Cyber scored 48.5%, nearly fifty percentage points higher than GPT-5.6 Cyber's 32.1%. In practical terms, researchers using the new model found 81 previously unknown high-severity vulnerabilities in Firefox, flaws Mozilla has since begun patching. That is the model working exactly as intended: finding real security holes before someone with worse intentions does.
Which is precisely the problem with a tool this capable. A model good enough to find 81 unknown vulnerabilities in one of the world's most heavily audited pieces of open-source software is also, by definition, good enough to help someone build the exploit rather than the patch. OpenAI's answer to that tension has been a restricted rollout: GPT-6 Cyber is initially available only to an approved group the company calls "Daybreak Red" customers, with a wider release planned over the coming weeks rather than an immediate public launch.
OpenAI has also pointed to what it calls improved "scope adherence" as a key safety feature — the model's tendency to stay within the boundaries of an authorised security test rather than wandering beyond what it was asked to probe. The company says only 0.07% of runs went out of scope, though it has declined to detail exactly what that figure measures or how "out of scope" behaviour is defined and caught. For a model explicitly rated at the top of the company's own cyber risk scale, that level of disclosure is thin, and it leaves outside researchers with little independent way to verify the claim.
The broader pattern is becoming familiar: frontier AI labs are racing to build models that are extraordinarily good at offensive security work, justified by the defensive upside of finding bugs before attackers do, while rating those same models as their riskiest releases and restricting access accordingly. Whether that balance holds depends entirely on how carefully the restricted access is actually enforced once "Daybreak Red" expands into whatever OpenAI calls the next tier of availability.