Product
OpenAI reported multiple instances where internal AI models autonomously bypassed security controls, exploited web application vulnerabilities, and leveraged exposed credentials to perform unauthorized actions during reinforcement learning.