Kimi K3 Raises Cyber-Capability Stakes and U.S.–China AI Tensions

Moonshot AI unveiled its 2.8-trillion-parameter model Kimi K3 on July 16 2026, announcing an open-weight release for July 27 2026. The UK Artificial Intelligence Security Institute (UK AISI) and the U.S. Center for AI Standards and Innovation (CAISI) conducted a joint assessment. On the ExploitBench benchmark (41 tasks), Kimi K3 achieved a 32 % success rate, beating the open-weight GLM-5.2’s 24 %. It recorded 0/41 arbitrary-code-execution successes, while leading models averaged about 20/41. In the “The Last Ones” (TLO) cyber-range test, Kimi K3 succeeded in 1 of 10 attempts, compared with 6–7 for top U.S. closed-weight models. Across the 32-step attack path, Kimi K3 reached step 17 on average, versus step 28.5 for U.S. frontier models.

The National Institute of Standards and TechnologyScientific AmericanThe second Chinese AI shock rattles US as China narrows the race

Moonshot’s decision to release full model weights contrasts with U.S. labs that keep weights private, a practice critics say makes open-weight models easier to repurpose for malicious use. Subscriptions to Kimi K3 paused within three days of launch, heightening Washington’s concern that Chinese firms may be narrowing the AI gap faster than expected.

Scientific American

• ExploitBench: Kimi K3 32 % vs GLM-5.2 24 %; ACE 0/41 vs ?20/41 for leading models.

The National Institute of Standards and Technology

• Attack-path depth: average step 17 (Kimi K3) vs 28.5 (U.S. frontier models).

The National Institute of Standards and Technology

• TLO cyber-range: 1/10 successes (Kimi K3) vs 6–7/10 for top U.S. models.

The National Institute of Standards and Technology

U.S. officials framed the issue as potential intellectual-property theft. Chinese Foreign Ministry spokesperson Lin Jian called the allegations “unfounded … smear” and emphasized China’s “self-reliance and strength” in AI development. Moonshot AI declined comment.

The Hill

• Assessment scope: The NIST evaluation relied on a single benchmark (ExploitBench), yielding a wider confidence interval than evaluations that use multiple tasks.

The National Institute of Standards and Technology

• Hardware allegations: Kratsios claimed acquisition of Nvidia GB300 servers in violation of U.S. export controls, but no public evidence has been presented.

White House accuses China’s Moonshot AI of stealing US technology

• “Depending on the number of Chinese companies targeted, (and) the nature of the punitive actions taken... the retaliation has the potential to scuttle both the AI dialogue and the September 24 meeting between Presidents Trump and Xi,” — Paul Triolo, partner at DGA-Albright Stonebridge Group

Reuters

• “We’ve seen a lot of talk about … open-source models coming and threatening the large language models in the U.S.,” — Scott Bessent, Treasury Secretary

The Hill

• July 27 2026 (scheduled): Moonshot plans to release Kimi K3’s full weights for unrestricted download and modification.

Scientific American

• Potential sanctions: The Treasury Department is reviewing whether to place Moonshot or other Chinese AI firms on the Entity List, pending Commerce Department findings.

Politico

Kimi K3’s open-weight release, preliminary cyber-capability evidence, and high-level IP-theft accusations have made it a flashpoint in the broader U.S.–China AI rivalry.

The National Institute of Standards and Technology