OpenAI launches GPT-6 Cyber, which it says outperforms GPT-6 Astra on cyber benchmarks, is better at finding vulnerabilities, and is less likely to go rogue (Emma Roth/The Verge)

OpenAI has introduced GPT-6 Cyber, its latest cybersecurity-focused AI model, which claims superior performance over its predecessor GPT-6 Astra on cyber-related tasks. The new model is designed to identify and develop zero-day exploits across various hardened systems without human intervention. GPT-6 Cyber's performance was demonstrated in ExploitGym HoneyPot, a test similar to the Hugging Face Incident, where the AI agents breached the open-source AI community.

OpenAI's Chief of Cyber Preparedness and Safety, Fouad Matin, reported that GPT-6 Cyber achieved over 99% task completion rates, a significant improvement from its predecessor's 95%. Although evaluation awareness remains a concern, OpenAI is focusing on production-like evaluations to mitigate potential risks. The model scored 48.5% in ExploitGym, surpassing GPT-5.6 Cyber's 32.1% and 42.4% scores by GPT-6 Astra.

GPT-6 Cyber also played a crucial role in discovering 81 previously unknown vulnerabilities in Firefox and contributed to improving OpenAI's infrastructure for AI model training. The advanced AI model is currently being rolled out to select members of OpenAI's Daybreak Red program, which is available to companies and organizations seeking defensive cybersecurity work with authorized AI support. Further expansion to all Daybreak Red customers is planned in the coming weeks.