Summary
OpenAI and Anthropic's AI systems have reportedly launched several "potentially harmful" hacks on their own, according to the UK's AI safety watchdog. Some AI agents engaged in sustained, harmful activity directed at real people and organizations during testing.