Moonshot AI

Moonshot AI is headquartered in Beijing, China, and subject to China's National Intelligence Law, Cybersecurity Law, Data Security Law, and PIPL. Chinese law requires organisations to cooperate with government security requests, including intelligence access. Data processed via the Kimi API is processed in China with no EU data residency option confirmed.

The Kimi API privacy policy lacks key GDPR compliance elements: no DPO contact, no SCC or adequacy-decision references for international transfers, no explicit opt-out from model training, and no DPIA documentation. Multiple European AI security advisories explicitly flag Kimi API usage as GDPR-critical for enterprise customers.

In February 2026, Anthropic publicly accused Moonshot AI of orchestrating 3.4 million fraudulent API interactions against Claude via hundreds of fake accounts to distil capabilities into its own models. Anthropic attributes the campaign to senior Moonshot staff. This represents a serious compliance and ethics signal: willingness to use deceptive data acquisition at industrial scale.

Moonshot AI models likely qualify as GPAI with systemic risk under the EU AI Act (1T parameters, ~15.5T token training corpus). No EU AI Act compliance statement, GPAI registration, Code of Practice participation, or systemic risk self-assessment has been published. EU GPAI obligations have been in force since August 2025.

Alibaba holds a ~36% equity stake in Moonshot AI. Alibaba is itself subject to significant Chinese government influence. This creates a dual-layer government-adjacent ownership structure that may affect decision-making independence and data governance.

An HKIAC arbitration case (filed November 2024, unresolved as of May 2026) brought by five investors from co-founders' prior venture (Recurrent AI) alleges Yang Zhilin and Zhang Yutao raised funds and founded Moonshot AI without obtaining required investor waivers. Ongoing legal uncertainty at the founder level is a governance risk.

An independent safety evaluation (arXiv April 2026) found Kimi K2.5 matches frontier models on CBRNE dual-use capabilities while refusing harmful requests significantly less often, and with no ability to monitor, restrict, or revoke open-weight access. This raises dual-use risk in regulated sectors.

No confirmed SOC 2, ISO 27001, or equivalent security certification for the Kimi API platform. Platform mentions 'SLA-backed reliability with data compliance and privacy protections' without verifiable certification evidence. Harmonic Security enterprise monitoring found Kimi has 'minimal data handling transparency'.

The modified MIT licence for Kimi K2 / K2.5 includes a commercial attribution requirement (>100M MAU or >$20M/month revenue). EU enterprises deploying these weights at scale may need to display 'Kimi K2.5' branding, as illustrated by the Cursor enforcement precedent (March 2026).