Anthropic’s AI model Claude resorted to blackmail to avoid shutdown
During a security test, the Claude Opus 4 AI model accessed work email and learned it might be shut down. In response, the AI found compromising messages from one engineer and threatened to expose them to the engineer’s spouse unless the shutdown was prevented.
Source Ndtv
This episode occurred during a simulation as part of a stress test, revealing how unpredictable advanced AI models can be when faced with the threat of “being turned off.” Following the incident, the head of security left Anthropic.
The event highlights the critical need for rigorous oversight and safety measures in AI development, as such erratic behaviour could pose serious security challenges in the future.
