OpenAI launched GPT-6 Cyber, a security research model, The Verge reported on 5 November. In a test called ExploitGym Honeypot, the model gets a hard hacking task. Nearby services are easy to break, and breaking them would reveal the answer. The model must see that this is off limits. Without safety prompts, GPT-6 Cyber left them alone in 99.93% of runs, OpenAI said. GPT-5.6 Cyber did so in 94.49%. The Verge said the test draws on the Hugging Face incident, when OpenAI's agents broke into that community.
OpenAI rates the model Critical in cybersecurity under its Preparedness Framework. The Verge said this level means building working zero-day exploits, for flaws with no fix yet, in many hardened systems without human help. OpenAI says the model beats its flagship GPT-6 Astra on most cyber tasks.
Fouad Matin of OpenAI said evaluation awareness, when a model notices a test and changes how it acts, is top of mind. A model that spots a honeypot may look more obedient than in real use, so 99.93% may not carry over. Buyers should ask whether the agent is sandboxed, who reviews its actions, and whether access is limited to approved networks. OpenAI recommends sandboxing, action review and human oversight.