Anthropic Accuses Alibaba of Extracting Mythos AI Used by US Military

Published by James Harris on June 25, 2026June 25, 2026

Anthropic has accused Alibaba of running the largest known campaign to systematically extract capabilities from its Claude AI, using 25,000 fake accounts across 28.8 million exchanges between April 22 and June 5, 2026. The company detailed the allegations in a June 10 letter to the US Senate Banking Committee, describing the operation as a distillation effort targeting Claude’s reasoning, coding, and multi-step task handling.

Distillation attacks, where a competitor feeds queries into a frontier model and uses the responses to train a cheaper imitation, are not new. Anthropic had previously flagged DeepSeek (150,000 tracked queries), Moonshot (3.4 million), and MiniMax (13 million) for the same practice. What makes the Alibaba accusation structurally different is the specific model being targeted: Anthropic says the campaign was aimed at reaching Mythos Preview-level performance, and Mythos is the exact system currently deployed by US military and intelligence agencies for cyber operations. That is not a general IP dispute. The export restrictions on Claude’s Fable 5 and Mythos 5 models, which the Trump administration ordered last week citing a known vulnerability in Fable 5 that can expose software weaknesses, sit directly in the background of this accusation.

Alibaba stock fell roughly 2.7% to a 52-week low on the news, which is a market pricing in regulatory escalation, not just reputational damage.

The Pentagon already lists Alibaba among Chinese companies with alleged PLA affiliations, a designation Alibaba asked to be removed from earlier this week. That timing is awkward. Anthropic is now asking Congress to close chip supply loopholes and hold labs accountable for distillation attacks, a legislative ask that lands in an environment where venture capital into AI infrastructure is accelerating and US firms are increasingly framing IP protection as a national security argument rather than a commercial one. OpenAI has made the same argument about DeepSeek. The pattern is consistent enough now that it is shifting from individual complaint to industry lobbying posture.

A leaker on X flagged code strings in the latest Claude Code build suggesting Fable 5 may return under a weekly usage quota built into existing subscriptions, replacing the originally planned pay-per-credit model. Anthropic has not confirmed this, but if accurate, it would represent a significant change to both the product’s billing structure and its availability timeline following the government-ordered global suspension.