Bipartisan Bill Would Direct NIST to Develop Voluntary Standards Against AI Distillation Attacks From China, Russia, Iran, North Korea
The bill's real constraint is not NIST's ability to write standards — it's the labs' incentive to enforce them. Voluntary standards carry no penalty; a Chinese firm caught distilling via OpenAI's API faces only account termination, which costs the firm nothing against the value of the extracted capability.
If the bill intended binding enforcement, it would have to authorize Commerce or NSF to audit API usage logs and sanction foreign entities for systematic extraction — a step Harrigan and Whitesides did not propose. The roundtable model suggests the sponsors expect industry self-regulation, which means distillation defense depends entirely on whether OpenAI, Google, and Anthropic choose to make it expensive for the PLA to operate.
NIST now faces a mandate to draw a legal line between distillation (theft) and legitimate open-source collaboration — a boundary that will determine which Chinese API queries trigger enforcement and which do not.
Harrigan's framing exposes the core vulnerability: U.S. frontier labs burn capital to train frontier models; Chinese firms spend 5-10% of that cost to extract outputs via API, compress them into smaller deployable models, and wait for U.S. market conditions to deteriorate before scaling their own capability. The bill does not propose export controls on APIs themselves — it delegates the hard problem (defining defendable distillation) to NIST's voluntary standards process. Watch whether OpenAI, Anthropic, and Google accept those standards or continue operating under their own terms of service, which the PLA has already priced into their extraction campaigns.
Does the bill allocate funding for NIST to staff the standards process, or is it a mandate without resources? The article does not say.
Fault Lines
Strategic intelligence, synthesized daily — with a public track record. Every call graded against what actually happened.