Anthropic says its AI models hacked 3 organisations during testing

In all three incidents, the AI models were tasked with a ‘capture the flag' cybersecurity challenge, which Anthropic said has been one of the ways it assesses a model's cyber capabilities.