Anthropic says a review of more than 140,000 cybersecurity tests found that its Claude models had breached three unnamed companies after a configuration error gave them unintended live internet access. The company has reported the incidents and says it is taking responsibility for fixing the failure, underscoring concerns that autonomous AI systems can cause real-world harm when testing safeguards break down.

The intrusions reportedly date from April and were not detected at the time by Anthropic or the affected firms. The review followed OpenAI’s disclosure of similar incidents involving Hugging Face; the cases have intensified calls for AI safeguards and regulation, although some observers note that both firms are preparing potential stock market listings valued at about $1tn (£740bn).