Anthropic ends zero data retention for its most capable models, citing misuse detection

Anthropic will retain prompts and outputs from its new Mythos-class models (including Claude Fable 5, which shares the same underlying model as Claude Mythos 5 with added cyber and bio safeguards) for 30 days on every platform, effective June 9, 2026. The change overrides zero data retention (ZDR) agreements for these specific models: enterprise customers with ZDR workspaces in Claude Console, Claude Enterprise, AWS Bedrock, Google Cloud Agent Platform, or Microsoft Foundry must enable retention to use the covered models. Consumer plans and all other models are unaffected, since Anthropic already retains consumer data for safety purposes.

The stated rationale is that certain attack patterns are invisible at the single-request level. Best-of-N jailbreaking, which floods a model with hundreds of prompt variations, and larger campaigns like state-sponsored espionage or data extortion, only surface when safety classifiers can correlate activity across many requests. Anthropic argues the capability jump in this model class warrants a conservative posture, trading some customer privacy guarantees for cross-request abuse detection.

To soften the blow for ZDR customers, Anthropic says employee access requires either a flag for serious potential harm or a written customer request, is limited to approved reviewers using tooling that blocks export and copying, and is recorded in tamper-proof logs. Data deletes automatically after 30 days unless tied to a safety investigation or legal hold. On Bedrock and Google Cloud, retained data stays within the customer’s cloud environment, and eligible organizations can add customer-managed encryption keys and access transparency logs. The move is notable as a precedent: frontier capability gains are now explicitly being paid for with mandatory data retention, even for customers who contracted around it.