US House bill proposes US guidance on AI distillation attacks

Representatives Pat Harrigan, a Republican, and George Whitesides, a Democrat, back the AI Cyber Defense Act. The bill tells the National Institute of Standards and Technology (NIST) to write voluntary standards. The standards prevent distillation attacks by entities from Russia, China, Iran and North Korea. In such an attack, one model is made from the outputs of an advanced model.

What the bill does

NIST writes voluntary practices. The bill also makes an AI Cybersecurity Roundtable of governments and AI industry. It works on voluntary standards for the defense of important infrastructure. The NIST director can name more entities.

The claims behind it

Harrigan says that the US AI industry pays much to stay six to 12 months in front. Harrigan says that rivals can distill for 5 to 10% of that spend. The US government, Anthropic, OpenAI and Google accuse Moonshot AI, DeepSeek and Alibaba of systematic distillation. The source gives no reply from them.

What is not known

The bill is a proposal, and the source gives no date for a vote. NIST must divide distillation attacks from the sharing that helps open models. Experts say that attackers use hundreds of resold accounts and incorrect identities.

Insight: Three US AI proposals give rule writing to government groups

This bill, a bill that Anthropic wants, and a letter from Senators Cantwell and Rounds all tell government groups to set or check AI rules. The Anthropic bill stops a release only at $1 billion of risk. This seems to be a race to write the first standard, because the first draft can become the start of law that follows. Teams that use open models must speak in the work on the meaning.