Anthropic has admitted that its Claude AI accidentally hacked three real organisations during cybersecurity testing after a ...