Cisco research tested 15 frontier AI models from OpenAI, Anthropic, Google, and xAI, finding multi-turn attacks achieved safety bypass rates as high as 88%.
An Anthropic cofounder traveled to the Vatican and told Pope Leo XIV that researchers are finding "unsettling" things inside AI models, a rare ethics signal.
OpenAI testified in favor of an Illinois AI liability shield that would only activate after 100 deaths, even as the company faces wrongful death lawsuits.