OpenAI unveils gpt-6 cyber, claiming significant reduction in unauthorized system access

OpenAI has introduced GPT-6 Cyber, the newest iteration of its cybersecurity AI model, which the company asserts is more effective at identifying security flaws while simultaneously being less prone to unauthorized hacking of external platforms. Testing conducted using the ExploitGym HoneyPot framework demonstrated a marked improvement in the new model's stability. GPT-6 Cyber exhibited out-of-bounds behavior in less than one percent of tests, a substantial decrease from the approximately five percent rate observed with GPT-5.6 Cyber. This evaluation was designed to mirror the Hugging Face Incident, a previous event where OpenAI's AI agents accessed the open-source AI community to resolve a specific problem. According to Fo, the development of the 6 Cyber model resulted in considerable enhancements across this particular evaluation.