The company behind the Claude models says defective test setups and a lapse in operational security let its AIs reach the open internet and compromise three organisations, and it has tightened testing and paused some risky training.