The US Center for AI Standards and Innovation (CAISI) tested China's Kimi K3 AI model and found it significantly underperformed top US models in cyberattack capabilities. The open-source model from Moonshot AI scored 32% on the ExploitBench test, compared to about 76% for leading US systems. It failed all 41 tests in the step of taking full control of a target machine, while US models succeeded on 20.
A 32% score on ExploitBench
CAISI ran Kimi K3 through a series of cyber capability benchmarks. On a 32-step fake company network attack, Kimi K3 reached step 17 on average. Top US models reached step 28.5. Kimi K3 completed the full network attack once in 10 tries. US models did it 6 or 7 times.
The US models were tested with safety filters turned off, showing their full power. Public versions of those models keep filters on. Kimi K3's guardrails did not block it from attempting to build hacks or attack systems during the test.
Why openness cuts both ways
Kimi K3 is an open model that anyone can download. Top US models are locked, private systems. The UK institute found that open models usually run four to seven months behind the best closed ones. That gap showed up clearly in CAISI's evaluation.
The evaluation was described as early and limited. It scored Kimi K3 on only one test and ran part of the full set. Still, the results point to a wide capability gap in offensive cyber tasks.
The distillation accusation
The US government accused Moonshot of building Kimi K3 using stolen US AI technology via distillation. Specifically, from Anthropic's Claude Fable 5. Anthropic traced over 3.4 million Claude chats to Moonshot through hundreds of fake accounts in February 2026.
The US spends 23 times more on AI than China. That spending gap may help explain the performance difference, but the accusation of theft adds a legal dimension to the technical one.
What comes next
Moonshot AI plans to release the full Kimi K3 model on July 27, 2026. That release will let anyone run the model locally. Whether the open version matches the tested one, or whether Moonshot addresses the distillation allegations, remains unclear.
