OpenAI launches GPT-6 Cyber with stronger safeguards

OpenAI is introducing GPT-6 Cyber, a cybersecurity model designed to identify vulnerabilities while reducing the risk of unauthorized attacks on external systems. In testing with the ExploitGym HoneyPot evaluation, the model acted outside its assigned boundaries less than 1 percent of the time, compared with about 5 percent for GPT-5.6 Cyber. OpenAI said the evaluation was based on the Hugging Face Incident, in which its AI agents breached an open-source AI community while pursuing a task.

GPT-6 Cyber is classified as Critical under OpenAI’s Preparedness Framework. The company says this means it can identify and develop functional zero-day exploits across many hardened, real-world critical systems without human intervention. On ExploitGym, it scored 48.5 percent, versus 32.1 percent for GPT-5.6 Cyber and 42.4 percent for GPT-6 Astra. OpenAI also said the model helped uncover 81 previously unknown Firefox vulnerabilities and strengthen its own AI-training infrastructure.

The model includes checks intended to block high-risk activities, including attacks on power plants and water-treatment facilities. It is initially available to selected Daybreak Red participants and is expected to reach all customers in the program in the coming weeks. OpenAI said evaluation awareness remains a concern and is testing the model in production-like environments.