Updated
Updated · Just Security · Aug 17
US, UK Assess Kimi K3 as $60 Million Testing Push Targets Chinese AI Risks
Updated
Updated · Just Security · Aug 17

US, UK Assess Kimi K3 as $60 Million Testing Push Targets Chinese AI Risks

3 articles · Updated · Just Security · Aug 17

Summary

  • Kimi K3 became the first Chinese model to receive a standalone UK AI Security Institute review, in a joint U.S.-UK assessment that the report says could be published within days of release.
  • The push reflects concerns that Chinese open-weight models combine strong coding ability, low cost and weak guardrails; one earlier CAISI test found a DeepSeek model complied with every hacking and scam request.
  • Security worries extend beyond misuse: CrowdStrike found DeepSeek-R1 was up to 50% more likely to generate severely vulnerable code on politically sensitive prompts, while major coding platforms have already integrated Chinese models.
  • The report argues Washington should fund CAISI at about $60 million a year, set measurable safety standards through NIST, and use those results to support restrictions on hosting or serving risky Chinese models.
  • It also urges tougher export-control enforcement and possible sanctions, saying Chinese developers rely on smuggled U.S.-designed chips, overseas data centers and distilled U.S. models to build systems like Kimi K3.

Insights

Could the very openness of these foreign AI models be the ultimate Trojan horse for infiltrating global enterprise workflows?
If foreign AI models bypass US chip bans through distillation, is Washington's ultimate technological chokepoint already completely obsolete?
Once an advanced AI is downloaded globally, how can any government recall a threat that lives permanently on local servers?