ANTHROPIC says it has strengthened Claude AI’s defenses against model distillation, a technique in which attackers harvest transcripts and reasoning logs to train smaller, inferior models. The firm reports that their new safeguards have dramatically cut the amount of extractable reasoning data by about 90 to 99 percent, making distillation far less viable in practice.
Earlier measures already limited cross-model retrieval of reasoning logs and prevented changes to the model’s prior conversational context; the latest updates extend these protections to active API use and data handling.
Internal testing is used to illustrate the impact: models trained on full reasoning logs scored 80.4 percent on SWE-bench Verified programming tasks, while those trained only on summarized or restricted reasoning performed 71.8 percent, highlighting the value of full logs in training. Anthropic notes that the precise outcomes of distillation can vary depending on methodologies and model architectures.
New data-protection rules apply to API accounts created on or after 31 August 2026, prohibiting editing of previous conversations and restricting inter-model reasoning-log retrieval to the originating organisation, Claude Sonnet 5.5. A two-stage classifier is deployed to identify and block requests attempting to extract hidden reasoning, with ongoing updates to counter evolving evasion methods.
The company also warns that while false positives are rare (under 0.1%), legitimate requests may occasionally be blocked, and some integrations may need reconfiguration.