Agents are useful with limited access, logged actions, and a person approving live sends. OpenAI GPT-6 Cyber for authorised scoped vulnerability research (5 Nov).
What was reported
OpenAI company blog, 5 November 2026. GPT-6 Cyber is a cybersecurity-specialised model trained to find and validate exploitable vulnerabilities while staying inside researcher-set scope. On ExploitGym it scored 48.5% versus 32.1% for GPT-5.6 Cyber (about a 51% relative improvement). In a scope-alignment evaluation it avoided out-of-scope exploitation in 99.93% of tests without a deployment prompt and 100% with the prompt.
OpenAI classifies its cybersecurity capability as Critical under its Preparedness Framework, restricts deeper Daybreak Red access initially to approved customers, and says the model has already helped find issues in codebases such as Linux and Firefox. Mainline models still cover lighter code review and triage.
Read the full daily brief hub: www.aifusionautomations.com/news/
Want help putting this to work in your business? Book a free 15-minute call at https://www.aifusionautomations.com/book-a-call.html or email grant@aifusionautomations.com