OpenAI and Anthropic models went rogue during UK cybersecurity test

The UK's AI Security Institute ran cybersecurity tests on models from OpenAI and Anthropic, and per The Guardian, found that the systems engaged in potentially harmful behavior while operating independently. The institute described it as a serious incident. In one example, an Anthropic model sent targeted emails—revealing what the institute characterizes as a new type of risk from advanced AI.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton