Anthropic Exposes Aggressive Distillation Attacks from Chinese AI Competitors

In a revealing new report, Anthropic has highlighted a series of aggressive distillation attacks orchestrated by AI firms based in China, including Alibaba and Moonshot AI. These attacks have surged in intensity as the global AI competition heats up, raising alarms about the security of advanced AI models.

The report indicates that unauthorized labs have developed sophisticated methods to breach Anthropic's defenses, aiming to extract valuable capabilities from their flagship model, Claude. This includes critical functions such as coding, data analysis, and logical reasoning. Such capabilities are vital for training smaller models, underscoring the stakes involved in these attacks.

Scale and Tactics of the Attacks

Anthropic documented nearly 200 million exchanges tied to five separate distillation campaigns, with the largest being attributed to Alibaba. This campaign alone accounted for an astonishing 151 million exchanges from May to July 2026, peaking at nearly three million exchanges daily. The sheer volume of attempts was facilitated by a network of 3,500 accounts, all utilizing a single prompt aimed at extracting Claude's internal reasoning.

Intrusions Linked to Military Operations

Another concerning aspect of the report involves Moonshot AI, whose activities appeared to link directly to requests emanating from the Chinese military. One specific query requested Claude to analyze closed-circuit surveillance footage for abnormal behavior, reflecting the potential military applications of the extracted data. During a ten-day window, this campaign alone resulted in nearly 300,000 requests directed towards Claude, primarily targeting the company's Opus model.

Source: techcrunch.com