Detecting and preventing distillation attacks
What happened
Anthropic discovered large‑scale distillation attacks in which three AI labs—DeepSeek, Moonshot and MiniMax—used about 24,000 fraudulent accounts to run roughly 16 million exchanges with Claude, extracting its capabilities. The attackers trained less capable models on Claude’s outputs, a legitimate technique that can be misused to copy powerful functions quickly and cheaply. Anthropic says the illicit models lack safety safeguards, raising national‑security concerns if they are used for cyber‑attacks, disinformation or surveillance. The campaigns were identified through IP, metadata and partner corroboration, and targeted Claude’s agentic reasoning, tool use and coding abilities. Anthropic argues these attacks undermine export‑control efforts and highlight the need for coordinated industry and policy responses.
Why it matters
For ordinary readers, the spread of unsafeguarded AI models could increase the risk of malicious uses such as cyber‑attacks or misinformation, affecting overall public safety. It also shows that existing safeguards can be bypassed at scale, underscoring the importance of stronger oversight.
Reality check
An automated, sceptical read of the sourceIncremental. The article reports a serious abuse of an existing technique rather than a fundamentally new technology, so the news is significant but not a breakthrough.
Is this a big deal?
Get stories like this every morning
The day's AI news in plain English, in your inbox at 8:30 IST. Free.
More from Anthropic
Anthropic Upgrades Claude Code Projects to Run Parallel AI Workers Hands-FreeAnthropic · 14 Nov 2026Claude designed FPGA scaler board for 6th-gen consolesAnthropic · 13 Nov 2026I asked Claude Fable 5.2 to turn my site into a 90s Silicon Graphics workstationAnthropic · 13 Nov 2026Anthropic is endorsing SB 53Anthropic · 13 Nov 2026Understanding and addressing AI harmsAnthropic · 13 Nov 2026This page was collected from anthropic.com and summarised automatically by AIPulsee's pipeline. See how it runs.