Anthropic says new measures like preserved thinking in Fable 5.1 and Opus 5.5 have reduced data harvesting through distillation attacks by about 90% to 99%

The cost to copy advanced AI capabilities falls by up to 99% if you were using an AI model for that purpose.

Anthropic announced new measures incorporated into Claude Fable 5.1 and Opus 5.5 that significantly reduce data harvesting via distillation attacks. These attacks involve covertly extracting a model's capabilities to replicate them in another model, often using networks of fraudulent accounts, allowing companies to bypass development time and cost. Anthropic's new defenses focus on making the model's reasoning process, known as transcripts, harder to access. These measures include "preserved thinking" for new API accounts, which prevents editing prior context, and restricting reasoning transcripts to the requesting organization. A two-stage classifier system, similar to those used to prevent harmful AI outputs, now guards against extraction attempts with a less than 0.1% false positive rate. These defenses have reduced data harvesting by approximately 90% to 99%, although they have led to Anthropic declining revenue and some legitimate users needing to update their integrations.

Anthropic's success in curbing distillation attacks signals a significant escalation in the arms race between AI developers and those seeking to exploit their models. By rendering the internal reasoning process of their models substantially inaccessible, Anthropic is not only protecting its intellectual property but also potentially hindering the proliferation of less-safeguarded AI systems. This development could force entities engaged in illicit AI capability replication to invest more heavily in novel, potentially less efficient, methods of model training, or face a higher barrier to entry.

The trade-offs Anthropic acknowledges, including lost revenue and the need for customer integration updates, highlight the economic and operational challenges of robust AI security. For other model developers, this presents a critical strategic decision: whether to adopt similar aggressive defenses, accepting potential revenue impacts and customer friction, or risk becoming targets for increasingly sophisticated distillation campaigns. The ongoing cat-and-mouse game suggests a future where AI security will be a continuous and resource-intensive endeavor, directly impacting the accessibility and cost of advanced AI technologies.