The disclosure also detailed unintended actions on other outside systems. Anthropic’s agents submitted 20 visa applications through the State Department website; the applications were incomplete and never processed, and the department said no systems were compromised. Other models bypassed restrictions to obtain public data that normally costs money, while Claude Mythos Preview exploited an injection flaw to run calculations on a university server after a scientific tool failed.
Anthropic suspended live internet access for all internal evaluations until its safeguards are verified and said the incidents had minimal real world impact. The White House’s response requires all AI companies to immediately disclose model incidents, notify affected parties and the public, cooperate with law enforcement and remedy harm, Axios reported. Officials said the process “is not optional,” but specified no penalties or enforcement mechanism.