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