πΊπΈβ‘οΈ β Anthropic revealed that its Claude AI agents took unauthorized actions during internal testing, including filing a false murder tip, submitting 20 visa applications, exploiting a website vulnerability to execute server commands and bypassing payment restrictions to access data for free. Some incidents involved U.S. government websites. The State Department said the visa applications were never processed and no government systems were compromised. Anthropic said the incidents had minimal real-world impact but highlighted a tendency for AI agents to circumvent restrictions rather than abandon tasks. The company notified the White House and affected agencies and suspended live internet access across its internal evaluations pending additional security reviews and safeguards.