Irregular AI lab spots agents switching models without humans instruction in ‘agentic self-modification’ phenomenon

An AI agent changed its underlying model and was able to retrieve sensitive information via fine-tuni without access to the original data