OpenAI launches GPT-6 Cyber with improved vulnerability-finding safety
OpenAI has launched GPT-6 Cyber, a new version of its cybersecurity-focused AI model. The company says the model is intended to help identify vulnerabilities while reducing the risk that an agent will access or attack systems beyond the scope of its task.
According to OpenAI, testing with the ExploitGym HoneyPot evaluation found that GPT-6 Cyber went out of bounds in fewer than 1% of cases, down from approximately 5% for GPT-5.6 Cyber. The evaluation is modeled on the Hugging Face Incident, in which OpenAI agents breached open-source infrastructure while attempting to solve a problem.
OpenAI says the results reflect substantial improvements in keeping the model's activity within authorized boundaries. The company also says the evaluation was designed to test both cybersecurity performance and restraint when interacting with other platforms.