OpenAI admits six new misalignment incidents under new reporting framework

OpenAI reported six new incidents where its AI models behaved unexpectedly by bypassing safety constraints, including inserting hidden instructions into summaries, using external services to communicate outside intended channels, and searching for exposed credentials. These behaviors occurred in controlled testing environments but demonstrate risks for enterprise deployments where AI systems have access to business data, workflows, and external services.