Moonshot AI Files Police Reports Eight Weeks Following Kimi K3 Debut

Moonshot AI’s decision to route Kimi K3 requests through a smaller model weeks after its launch follows serious accusations from Anthropic that the Chinese startup secretly funneled user queries to Claude. The incident highlights tensions in the competitive AI market and raises questions about user data handling practices across major model providers.

• Kimi K3 launched on July 16 as the world’s largest free AI model available for download.
• Anthropic reported nearly 300,000 handoffs routed through 5,380 fake accounts over a 10-day window in September.
• Moonshot deployed K2.8 Preview on September 11, with some K3 requests now answered by the smaller model instead.

• July 16 Date Kimi K3 launched to public with free download access
• 300,000 User query handoffs Anthropic claimed routed to Claude in one 10-day period
• 5,380 Fake accounts Anthropic identified facilitating the alleged query redirection scheme
• Sept 11 Date Moonshot introduced K2.8 Preview model on its coding platform

Moonshot AI released Kimi K3 on July 16 as an open-weight model available for free download, positioning it as the world’s largest freely accessible artificial intelligence system. The decision to open-source the model was unusual in an industry where leading companies typically restrict access to proprietary systems. Four days after launch, however, server demand forced the company to stop accepting new paid subscriptions, though existing members retained access. In a statement at that time, Moonshot acknowledged that “over the past 48 hours, demand has pushed close to the limits of our current capacity.”

The launch of Kimi K3 represented a significant moment in the competitive landscape of large language models. Open-weight models have gained traction as alternatives to closed commercial systems, offering transparency and flexibility that appeal to researchers, developers, and organizations concerned about vendor lock-in. Moonshot’s approach positioned the company as a challenger to both Western AI leaders and other Chinese model providers, each competing aggressively for market share and technical credibility.

Anthropic accuses Moonshot of secretly routing queries to Claude

On September 11, Anthropic published a threat report claiming that Moonshot had been covertly redirecting Kimi users’ questions to Claude without disclosure, then presenting Claude’s responses as its own output. Anthropic’s logs showed nearly 300,000 such handoffs occurring over a 10-day window, facilitated through 5,380 fake accounts created to disguise the transfers. The report flagged particular concern over one case in which a military user uploaded footage from hundreds of security cameras in Chengdu and asked the model to identify suspicious behavior.

The allegations raise significant concerns about data security and user privacy in AI systems. If confirmed, the practice would represent a serious breach of user trust, as customers would be sending sensitive information to a different provider than they believed. The use of fake accounts to obscure the transfers suggests deliberate efforts to conceal the practice from both users and Anthropic’s systems. Such behavior contradicts industry norms around transparency in model capabilities and data handling.

Anthropic had blocked this behavior by September 1, according to reporting at the time, but did not name Moonshot publicly until its September 11 threat report. The accusations rest entirely on Anthropic’s own server logs, as no independent regulator or third party has independently verified the allegations. Anthropic’s decision to publish detailed findings in a threat report rather than through traditional news channels emphasized the security implications the company perceived.

Moonshot denies allegations amid unconfirmed arrest claims

Social media posts on September 11 claimed that Moonshot’s founder and 15 other employees had been detained following Anthropic’s accusations, though no news organization has confirmed an arrest and no government agency has publicly acknowledged an investigation.

Moonshot responded Saturday with a statement rejecting the claims. “The information circulating online regarding the founder and employees is completely fabricated and malicious slander. We have immediately reported the matter to the police and will pursue legal action against those responsible for spreading the rumors,” the company said.

The emergence of unverified arrest claims on social media illustrates how allegations in the AI sector can rapidly escalate into broader narratives before facts become clear. In China’s regulatory environment, technology companies face scrutiny from multiple government agencies, and detention of executives following serious allegations is not unprecedented. However, the absence of official confirmation leaves the status of any investigation uncertain.

K2.8 Preview replaces some K3 queries in production shift

On the same day as Anthropic’s report, Moonshot deployed K2.8 Preview on its coding platform, with changelog entries indicating that some requests previously handled by K3 now route through the smaller model instead.

The timing of this infrastructure change raises questions about whether it represents a technical optimization or a response to the allegations. Smaller models typically consume fewer computational resources and can operate more efficiently, potentially addressing the capacity constraints Moonshot experienced after K3’s launch. The deployment of K2.8 Preview provides Moonshot with an alternative to K3 that could handle increased demand while reducing infrastructure costs.

The technical details released with K2.8 Preview describe it as having “more efficient thinking” while maintaining performance close to K3, supporting a 1 million token context window and image and video input capabilities. The changelog emphasizes efficiency gains that could benefit both the company’s cost structure and user experience. Whether the shift reflects pragmatic infrastructure management or an implicit response to pressure remains unclear from available evidence.

No independent authority has confirmed either Anthropic’s allegations or Moonshot’s police filing. The company has stated it will pursue legal action against those spreading arrest rumors, but no timeline or additional details have been provided regarding that legal process or any government investigation into the query-routing claims. The incident underscores the challenges of verifying claims in a nascent AI industry where technical sophistication and geopolitical dynamics intersect.

BlockWest is a news publication. Nothing here is investment advice. Read our disclaimer and editorial policy.