OpenAI releases GPT-6 Cyber system card, focusing on cybersecurity capabilities and alignment performance
GPT-6 Cyber System Card
AI BriefingOpenAI released the GPT-6 Cyber system card, describing how reinforcement learning trains the model in cybersecurity capabilities such as reproducing crashes, developing working exploits, and searching codebases for vulnerabilities. In the ExploitGym Honeypot evaluation, its unauthorized exploitation rate without deployment safety instructions was 0.07%, far below GPT-5.6 Sol's 49.5%. Its violation rate for compliance with Auto-Review rejections was 0%, matching GPT-6 Astra.
Source: OpenAI: Deployment Safety and System Cards · deploymentsafety.openai.com
Event follow-up1
View all event coverage- OpenAI releases GPT-6 Cyber system card, focusing on cybersecurity capabilities and alignment performanceOpenAI: Deployment Safety and System Cards · 11 days ago