TechnologyTracking
Anthropic anti-distillation: principle and limits?
Anthropic claims significant progress in reducing the effectiveness of distillation attacks on Claude, but what is the specific technical principle behind this? Can it fundamentally prevent model distillation, or does it merely increase the cost of attacks?
Watching: Anthropic后续发布的技术细节或独立研究对防御效果的验证。
Current read
Anthropic reports a 90-99% reduction in reasoning-token extraction from distillation attacks, but the technical mechanism is not specified, so it is unclear whether it fundamentally blocks distillation or merely raises costs.
Available evidence shows that Anthropic published a blog on October 12, 2026, claiming its new defenses reduce the amount of reasoning tokens extracted by distillers by approximately 90-99% (i.e., to 1/10-1/100 of original levels), and mentions a mechanism of 'two-level classifiers plus a retained-thinking mechanism.' Additionally, a 154-page report based on 1.9 billion API interactions from seven labs involves detecting anomalous calls. However, these sources do not detail the specific technical principles (e.g., whether they limit reasoning exposure or detect proxy networks), nor do they provide independent verification. Thus, the evidence only confirms official claims of significant effectiveness but cannot determine whether the defense fundamentally blocks distillation or merely increases attack costs.
Key developments
- The assessment representation was corrected; evidence and earlier records are retained.
How we read it
Anthropic's official blog claims a 90-99% reduction in reasoning-token extraction, but the specific technical mechanism is not described, so only the official claim is confirmed, not its fundamental nature.
A 154-page report based on 1.9 billion API interactions from seven labs involves detecting anomalous calls, suggesting detection measures, but no independent verification or detailed mechanism is provided.
Earlier updates
- — Public tracking started; no independent probability estimate yet.; Starting public context: Anthropic says it made significant progress reducing effectiveness of distillation attacks on Claude
- — Anthropic reports a 90-99% reduction in reasoning-token extraction from distillation attacks, but the technical mechanism is not specified, so it is unclear whether it fundamentally blocks distillation or merely raises costs.; Available evidence shows that Anthropic published a blog on October 12, 2026, claiming its new defenses reduce the amount of reasoning tokens extracted by distillers by approximately 90-99% (i.e., to 1/10-1/100 of original levels), and mentions a mechanism of 'two-level classifiers plus a retained-thinking mechanism.' Additionally, a 154-page report based on 1.9 billion API interactions from seven labs involves detecting anomalous calls. However, these sources do not detail the specific technical principles (e.g., whether they limit reasoning exposure or detect proxy networks), nor do they provide independent verification. Thus, the evidence only confirms official claims of significant effectiveness but cannot determine whether the defense fundamentally blocks distillation or merely increases attack costs.