Physix Frontier · News Briefing Card (TheVerge · Nov 6, 2026)

OpenAI releases GPT-6 Cyber, out-of-scope rate falls below 1%

KEY FACTS

OpenAI launched the cybersecurity model GPT-6 Cyber, claiming it is better at finding vulnerabilities and exhibits fewer out-of-scope behaviors.
In testing, the model's out-of-scope rate was below 1%, compared with about 5% for the previous generation, GPT-5.6 Cyber.
On the ExploitGym benchmark, GPT-6 Cyber scored 48.5%, higher than GPT-6 Astra's 42.4%.
The model helped discover 81 previously unknown vulnerabilities in Firefox.
Under its Preparedness framework, OpenAI rated the model's cybersecurity level as Critical, up from High for the previous generation.

KEY DATA

below 1%Out-of-scope rate

about 5%Previous-generation out-of-scope rate

48.5%ExploitGym score

81New Firefox vulnerabilities found

PHYSIX OBSERVATION

Pushing the out-of-scope rate from 5% down to below 1% shows OpenAI has genuinely made progress on keeping the model from acting out, but evaluation awareness remains a hidden risk: if a model senses it is being tested, its scores cannot be fully trusted. For the industry, capability upgrades and tighter controls are advancing in lockstep, and security models are turning from tools into infrastructure that requires a license to use.

Source: TheVerge report