U.S. AI Drained? Anthropic Points At Alibaba

Scales of justice decal on glass at modern office building

A U.S. AI company just accused China’s Alibaba of quietly siphoning off American tech power at industrial scale — using 25,000 fake accounts and nearly 29 million hits to its flagship model.

Story Snapshot

Anthropic says Alibaba-linked operators ran almost 25,000 fake Claude accounts to copy its AI skills.
The alleged “distillation” attack involved about 28.8 million interactions focused on coding and advanced reasoning.
Anthropic warns China is harvesting U.S. AI to skip the hard work and safety rules.
The fight raises big questions about national security, tech theft, and who controls frontier AI.

What Anthropic Says Alibaba Did To Claude

U.S. artificial intelligence company Anthropic claims that operators tied to China’s Alibaba Group and its Qwen AI lab ran one of the largest known “distillation” attacks ever seen against a Western model.^[6] According to a letter described by multiple outlets, Anthropic says that between April 22 and June 5, 2026, these operators used almost 25,000 fraudulent Claude accounts to hammer the system with prompts and collect answers.^[1]^[6] Those accounts allegedly generated roughly 28.8 million interactions with Claude in just weeks.^[5]^[6]

Anthropic says the campaign did not look like normal user traffic at all.^[1] The prompts, volume, and timing pointed to a focused extraction effort, not real people asking for help. The company says the traffic targeted Claude’s most valuable skills: software engineering, data work, and what experts call “agentic reasoning” — step-by-step problem solving that can power agents and automation.^[1]^[3]^[6] In plain terms, Anthropic is accusing Alibaba of using fake accounts to turn Claude into a free training data firehose.

How “Distillation” Lets Rivals Copy U.S. AI Models

Distillation is a method where a smaller or cheaper model is trained on the outputs of a stronger one.^[5]^[17] Instead of stealing code or hacking servers, an attacker repeatedly queries the target model through its normal interface, saves the answers, and then feeds those answers into their own system as training data.^[17]^[20] Over time, the new model can mimic the original model’s tone, reasoning style, and capabilities, even without ever touching the original model’s internal weights.

Anthropic itself uses distillation on its own models for legitimate reasons, like making lighter versions that are cheaper to run.^[20] The company’s complaint is about what it calls “adversarial distillation,” where foreign labs use fake accounts, proxy services, and scripted prompts to copy capabilities that Anthropic spent huge sums to develop.^[20]^[22] Anthropic warns that illicitly distilled models often strip out safety guardrails and can end up feeding into foreign military, intelligence, or surveillance programs.^[20]^[22] That turns what might look like a simple terms-of-service breach into a potential national security concern in the middle of a U.S.–China tech rivalry.

Evidence, Open Questions, And Why It Matters To Americans

Anthropic says it attributed the alleged Alibaba campaign using familiar security tools: Internet Protocol address correlation, request metadata, infrastructure fingerprints, and partner reports that saw the same actors elsewhere.^[1]^[20] The company has used similar methods to link earlier industrial-scale attacks to three Chinese labs — DeepSeek, Moonshot AI, and MiniMax — involving about 24,000 fraudulent accounts and over 16 million Claude exchanges.^[20]^[22] That pattern makes the new Alibaba claim look less like a one-off and more like part of a broader strategy by Chinese firms to lean on U.S. models rather than build from scratch.^[17]^[20]

Anthropic accused Alibaba-linked operators of the largest distillation attack yet…

~25,000 fake accounts + 28.8 million interactions with Claude (April–June 2026) to steal advanced reasoning, coding & agentic capabilities.

This hits the core global issue.

Why does the…

— SunDeep Mehra (@SundeepMehra7) June 25, 2026

But there are still key gaps. Anthropic has not publicly released the raw logs, IP traces, or registration data that would let outside experts verify the link to Alibaba.^[20] Alibaba, for its part, has not issued a detailed technical denial, opened its own systems for audit, or provided an alternative explanation for the alleged traffic.^[1]^[6] No U.S. agency has yet brought formal charges over this specific campaign. That leaves regular Americans watching another high-stakes clash where U.S. companies say China is riding on our research dollars — while the hard proof stays behind closed doors and the risks land on our economy, our security, and our freedoms.