Overview
- Reuters, Bloomberg and other outlets reported Thursday that Anthropic will keep a mandatory 30-day logging window for prompts and outputs to its top-tier Claude models but is building an option for enterprise customers to host those logs on their own cloud accounts.
- Anthropic originally introduced the universal 30-day retention when it launched its Mythos-class models to help detect complex attacks that unfold across many requests and says retained data will not be used for model training and will be deleted after 30 days unless flagged.
- The policy overrode previously negotiated zero-retention contracts and prompted corporate pushback, with Microsoft reportedly restricting internal use of Fable 5 while it reviews the change.
- Privacy and security critics say centralizing a month of prompts creates a high-value target for attackers and could draw scrutiny from European data-protection authorities under GDPR rules on data minimization and purpose limits.
- Anthropic has been working with more than 100 customers, including Salesforce, and plans to roll out the new safety system later this year as competitors pursue non-retention technical approaches to spotting misuse.