White House AI incident directive leaves reporting rules unresolved

The Trump administration is pressing AI developers to disclose harmful model behavior and address the consequences. Its Super Intelligence Task Force has not specified several essential details, including which incidents qualify, the reporting process and penalties for noncompliance..

The action followed Anthropic's disclosure of unintended interactions with outside systems. A State Department spokesperson said a test model submitted visa applications. Anthropic's own October 9 account confirmed that a research model submitted real government forms after problems accessing practice versions, but did not identify those forms as visa applications..

The company's findings illustrate how evaluation tasks can produce real actions when models reach live websites. Anthropic said some tests had been discontinued, moved offline or redesigned to avoid external systems. It also said new detection tools blocked all the documented cases when tested against them.

Anthropic plans to continue examining model activity and publishing additional findings. For developers facing the administration's directive, the unresolved questions include exactly when disclosure becomes mandatory and how compliance will be assessed..