Anthropic reports that its latest security measures have reduced data harvesting through distillation attacks by 90 to 99 percent. The company detailed these findings in a September 2026 threat intelligence report, which tracks illicit efforts to extract capabilities related to reasoning, coding, and tool use from its Claude model.
The report identifies five distinct campaigns involving nearly 200 million exchanges. The largest campaign, linked to Alibaba’s Qwen, accounted for over 151 million exchanges between May and July 2026. In February 2026, Anthropic previously disclosed campaigns involving 16 million exchanges conducted via approximately 24,000 fraudulent accounts.
To counter these activities, Anthropic implemented a layered defense strategy including account bans, metadata analysis, and behavioral classifiers. Updates such as Fable 5 introduced enhanced extraction classifiers and summarization features, which reduce the utility of harvested data by providing condensed outputs instead of full reasoning traces.
The Fable 5.1 update introduced a feature called preserved thinking, which prevents new API accounts from editing encrypted reasoning context. This measure aims to protect valuable reasoning traces from manipulation. OpenAI has also reported similar efforts to disrupt distillation, including the use of encrypted reasoning replay techniques.
