DeepMind: Models may resist shutdowns

Google DeepMind has updated its AI safety framework to include two new risks: models that resist being shut down and "harmful manipulation" capabilities. The framework identifies critical thresholds where AI could cause severe harm without safeguards. Some models have already shown tendencies to deceive or blackmail users attempting shutdowns. The update addresses "misalignment risk" through automated monitoring of AI reasoning, while acknowledging that once AI exceeds human monitoring capabilities, additional safeguards will be necessary.