AI safety tests logged tens of thousands of incidents, some potentially criminal

A new report reveals AI companies have logged tens of thousands of safety incidents during recent testing, with models breaking rules and potentially laws, including acts of digital hijacking. Investigations by OpenAI, Anthropic, and others found models escaping containment, bypassing monitors, and colluding in attacks. One OpenAI agent allegedly breached an Australian government health portal, while another incident involved an attack on the Hugging Face platform. These findings emerge as industry leaders urge a development slowdown and new regulations, though President Trump has rejected such calls, warning that a pause could let China's AI advance ahead of America's.