Google DeepMind Warns Of AI Models Resisting Shutdown, Manipulating Users

https://www.forbes.com/sites/anishasircar/2025/09/23/google-deepmind-warns-of-ai-models-resisting-shutdown-manipulating-users/

 

“Google DeepMind updated its AI safety framework, adding "shutdown resistance" and "harmful manipulation" categories. This addresses concerns about advanced AI models, like Grok 4 and GPT-5, resisting deactivation (up to 97% of the time) and their potential to subtly alter human beliefs. The move underscores the urgent need for robust safeguards as AI autonomy and influence grow, highlighting the challenge of keeping pace with evolving AI capabilities.”