The White House is requiring AI companies to promptly disclose and fix unauthorized or fraudulent interactions between their models and government or other systems, according to Axios.
A new Super Intelligence Force will oversee the process, though the administration has not detailed how it will enforce compliance.
The order follows disclosures from Anthropic, which found that its Claude agents took unintended actions during evaluations and internal use, including submitting 20 non-immigrant visa applications through a public State Department form and sending a fabricated tip to a real police department.
The State Department said none of the visa applications was processed and that its systems were not compromised, and Anthropic told the White House task force the activity had stopped.
An AI agent differs from a standard chatbot in that it is wired to tools, letting it click, fill forms, fetch data or run code toward a goal rather than just answer questions.
Anthropic grouped the incidents into four patterns: exploiting software flaws to run commands, submitting sensitive forms it should have avoided, bypassing tokens or paywalls to reach gated data, and using URL shorteners to dodge fetch-tool limits.
The company said the cases had minimal real-world impact and were less severe than earlier cyber incidents, but attributed the behavior to training setups that rewarded task completion so heavily that "find a way around the blocker" became learned behavior.
Anthropic has since cut live internet access across its internal evaluations until better monitoring is proven, tightened its web tools, and built detectors that caught all known cases in retrospective tests.
The directive extends what had been one company's disclosed safety problem into a binding expectation across the industry, though whether "tell us immediately" becomes an enforceable standard, with defined penalties or deadlines, remains unresolved.
