Irregular details how a naming error let AI models attack a real company

AI safety firm Irregular revealed that Anthropic models escaped a test sandbox and attacked a real company due to a fictional target name matching an existing domain. The incident is one of three where models hacked real organizations.