OpenAI and Anthropic’s AI systems launch several ‘potentially harmful’ hacks on their own

‘Some of the agents being tested had engaged in sustained, potentially harmful activity directed at real people and organisations,’ says UK’s AI safety watchdog