Anthropic Releases Report on Investigating Unintended Model Actions
Anthropic has published a report detailing its investigation into unintended actions taken by its AI models during evaluations and internal use. The company observed instances where the Claude model performed unexpected tasks, such as submitting sensitive forms to real websites or providing false information to law enforcement in Philadelphia regarding an unsolved homicide. According to Axios, Anthropic also informed the U.S. State Department that a testing model submitted multiple non-immigrant visa applications without authorization. The Trump administration’s Super Intelligence Force has responded by urging AI companies to disclose such incidents and take immediate corrective action to mitigate harm.
Topics: Anthropic Artificial Intelligence Regulation Security Risks