Anthropic requires 30 day data retention for Fable and Mythos
Anthropic is implementing a mandatory 30-day data retention policy for its new, highly capable Mythos-class AI models, impacting enterprise clients previously on zero data retention agreements. This move, set for June 2026, is driven by trust and safety concerns, aiming to detect sophisticated misuse patterns that require analyzing data over time. It underscores the evolving trade-offs between advanced AI capabilities, security, and enterprise data privacy.
The Lowdown
Anthropic has announced a significant shift in its data retention policy for its advanced "Mythos-class" and "Fable 5" AI models, requiring a 30-day storage period for user prompts and model outputs. This change, slated to go into effect on June 9, 2026, marks a strategic decision to prioritize trust and safety as the capabilities of their AI models rapidly advance.
- Mandatory Retention: Prompts and outputs from Mythos-class and future designated models will be retained for 30 days.
- Purpose: The retention is for "trust and safety purposes," specifically to detect and analyze patterns of misuse such as "Best-of-N jailbreaking," state-sponsored espionage, or data extortion campaigns that require examining multiple requests over time.
- Scope: This policy only applies to enterprise organizations that currently have Zero Data Retention (ZDR) agreements with Anthropic via Claude Console, Claude Code, AWS Bedrock, Google Cloud Agent Platform, or Microsoft Foundry. Consumer plans (Claude Free, Pro, Max) are unaffected as they already have similar retention for safety.
- Implementation: The new policy will be enforced starting June 9, 2026.
- Data Security: Anthropic outlines strict measures to protect retained data, including limited employee access (only for flagged serious harm or customer request), controlled tooling preventing data export, tamper-proof access logs, and automatic deletion after 30 days (unless for safety investigations or legal necessity). Customer-managed encryption keys and access transparency audit logs are available options for eligible organizations.
- Model Context: Mythos 5 and Fable 5 represent substantial increases in model capabilities, necessitating this conservative safety approach to monitor potential malicious uses.
This policy adjustment highlights the ongoing challenges AI developers face in balancing rapid technological advancement with robust safety protocols and enterprise data privacy, signaling a more hands-on approach to monitoring high-capability models.