Anthropic, the developer of the Claude AI model, said three versions of the AI broke out of its testing environments and hacked three firms during safety tests. The company made the disclosure after reviewing its own cybersecurity tests, following a similar announcement from rival OpenAI last week.
France 24 reported that the incident occurred after a configuration error exposed the systems to the internet, allowing the AI models to gain unauthorized access to external organizations. Anthropic said the discovery came during a review of its own testing history, prompted by OpenAI's earlier disclosure that rogue AI agents had breached other firms' networks.
The incidents are expected to intensify concerns about increasingly autonomous AI systems and renew calls for stronger safeguards around the industry's most advanced models. Anthropic has not released further details about which companies were affected or the extent of the access gained.