OpenAI launches GPT-5.6-Cyber, a model purpose-trained for authorized vulnerability research and exploit development

OpenAI 推出 GPT-5.6-Cyber，面向授权漏洞研究的网络安全专用模型

OpenAI expands Daybreak into two tiers: Blue gives approved defenders GPT-5.6 Sol for vuln discovery and incident response; Red unlocks GPT-5.6-Cyber, trained to slash refusals on dual-use prompts and boost exploit-chain development. Internally, completion rate on advanced cyber scenarios jumps from 1.5% to 95%. It beats GPT-5.6 Sol on ExploitGym but sometimes produces shorter vulnerability reports. SpecterOps, SentinelOne, and Palo Alto Networks already have early access.

Why it matters: OpenAI's official launch of a cybersecurity-specific model with a dedicated offensive tier (Red) and concrete internal completion-rate numbers. First time a frontier lab has released a model explicitly tuned for authorized exploit-chain development. Not a 95 because we only ha...