OpenAI Launches GPT-5.6-Cyber for Security, Uncovering 400+ Kernel Vulnerabilities

新智元 · wechat · 2026-08-11

OpenAI is expanding its 'Daybreak' initiative by launching GPT-5.6-Cyber, a model specifically designed for cybersecurity tasks to shift the security paradigm.

Demonstrating strong 'cyber detective' capabilities in real-world codebases, the model has discovered over 400 privilege escalation vulnerabilities in a major OS kernel and successfully chained two vulnerabilities in the Chrome V8 engine to achieve sandbox escape. It also found a complete local privilege escalation chain in a mainstream mobile OS.

To balance offense and defense, OpenAI divided the program into Blue and Red channels: the Blue channel provides defenders with a general model without systemic guardrails hindering legitimate defense; the Red channel offers GPT-5.6-Cyber, which fulfills 95% of advanced security requests compared to 1.5% from general models, freeing researchers from fighting model refusals.

Related event: OpenAI launches GPT-5.6-Cyber, expands Daybreak initiative(12 posts)→

Original post →

More from Models

Models channel →