CNAS presses NIST’s CAISI for 72-hour Chinese AI reviews
The useful question is whether Commerce can publish warnings before enterprises and researchers normalize the models.
TL;DR
Inside Cybersecurity reports that the Center for a New American Security urged Commerce to direct NIST’s Center for AI Standards and Innovation to publish national security assessments of Chinese frontier AI models within 72 hours of release, with classified or confidential material to agencies and allies within 96 hours. CNAS wants $10 million to $20 million annually for the mission; a House fiscal 2027 NIST bill includes $15 million. The point is speed: warnings after adoption mostly document the damage.
CNAS is asking Commerce to turn CAISI’s China model work from a research cadence into an alert cadence. Inside Cybersecurity reports that the June 12 report urges NIST’s Center for AI Standards and Innovation to publish dedicated national security assessments of Chinese frontier models within 72 hours of release, including agent-hijacking, jailbreak and censorship testing, with classified or confidential elements delivered internally within 96 hours.
The ask is not coming from nowhere. NIST says CAISI leads evaluations of U.S. and adversary AI systems, foreign AI adoption, and potential vulnerabilities or malign foreign influence in adversary systems (https://www.nist.gov/caisi). CAISI has already published assessments of DeepSeek and Moonshot models. Its September 2025 DeepSeek review found security and censorship risks, including agent-hijacking and jailbreak weaknesses that NIST said may affect developers, consumers and U.S. national security (https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks).
The hard part is the clock. A 72-hour assessment will be preliminary by definition, and CNAS concedes sensitive findings may need classified or confidential channels. But preliminary is the whole point if U.S. enterprises and researchers are already testing open-weight PRC models before an official evaluation lands. CNAS also wants $10 million to $20 million a year for CAISI’s China evaluation mission; Inside Cybersecurity says a House fiscal 2027 spending bill advanced May 13 includes $15 million. Congress can buy capacity. Commerce still has to decide whether CAISI is publishing postmortems or warnings.
Published ·Deep Fathom