
The White House says Moonshot AI used Anthropic’s Fable 5 in a large-scale distillation effort, while Anthropic has reported fraudulent account activity involving Kimi models.
White House science and technology chief Michael Kratsios also said Moonshot developed an internal platform that switched between access methods to avoid detection and acquired Nvidia GB300-equipped servers, including systems accessed in Thailand. The allegations do not amount to independently published proof of the full claim, but they affect how the planned open-weight release may be viewed.
What the White House Accused Moonshot AI of Doing
According to Kratsios, Moonshot AI distilled Anthropic’s Fable 5 for development of the K3 model. He said the company ran large-scale distillation against U.S. models and built a sophisticated internal platform that could quickly change how it accessed them.
The accusation includes a second infrastructure claim. Kratsios said Moonshot obtained GB300-equipped servers and accessed GB300 systems in Thailand, likely to train its AI models.
The U.S. government says legitimate distillation can help create smaller, more efficient models. It draws a distinction between that process and covert industrial-scale extraction of proprietary U.S. technology.
How Anthropic Detected the Alleged Campaigns
Anthropic’s February 23 analysis provides more detail about the activity attributed to Moonshot. Anthropic said Moonshot’s Kimi models employed “hundreds of fraudulent accounts spanning multiple access pathways.”
Varied account types reportedly made the campaign harder to detect as one coordinated operation. Anthropic said it attributed the campaign through request metadata that matched public profiles of senior Moonshot staff. The company also reported that Moonshot later attempted to extract and reconstruct Claude’s reasoning traces through a more targeted approach.
Anthropic’s wider complaint also named DeepSeek and MiniMax. It accused the three companies of running “industrial-scale campaigns” to “illicitly extract Claude’s capabilities.”
What Kimi K3’s Cyber Evaluation Shows
A preliminary cyber-capability assessment from the UK AI Security Institute and the US Center for AI Standards and Innovation provides a measured counterpoint to the launch benchmarks. The assessment reported that “Kimi K3 achieved ACE on 0/41 samples, whereas the most cyber-capable models achieved ACE on 20/41 samples on average.”
ACE means arbitrary code execution. The assessment described it as the highest-severity outcome in exploit development because it allows an attacker to hijack a target.
The same evaluation said Kimi K3 scored 32% on ExploitBench, compared with 24% for GLM-5.2, which the evaluators identified as the most cyber-capable open-weight model as of June 2026. On the Cyber Range: The Last Ones evaluation, a simulated 32-step corporate network attack spanning four subnets and about 20 hosts, K3 performed significantly below leading U.S. cyber-capable models.
Why the Allegations Matter for Open-Weight Releases
According to the report, Moonshot’s Kimi K3 was released last week with scores at or near the U.S. frontier, with the full model scheduled to have its weights published on July 27. If the distillation claim is substantiated, the release could be viewed as a wider distribution of capabilities allegedly extracted from proprietary U.S. models.
The documented account activity and access methods provide some evidence, but the training and hardware-use questions remain unverified. Many predictions also expect the U.S. to use the distillation claim against open Chinese models.
What to Do Now
- Track the July 27 weight release. Review the published model files, accompanying documentation, license terms, and technical notes when they become available.
- Treat the distillation claim as an allegation. Use the documented account activity and access methods, while leaving the unverified training and hardware-use questions open until more evidence appears.
FAQ
Related coverage
- White House Hack-Back Program Framework
- Moonshot AI’s Kimi K3 Beats Opus 4.8 as the Largest Open Model Ever – BizScoreAI
- White House AI Voluntary Cybersecurity Framework Meeting
Sources
See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.