Skip to content

Anthropic

Anthropic plans to loosen data retention on top models

2 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Anthropic plans to give business customers more control over their data on its most capable models, Bloomberg reported on August 20. The change would ease a rule that kept those customers’ prompts and outputs for 30 days, and the report says a new safety system is due later this year.

Anthropic set the rule in June, alongside the launch of Claude Fable 5 and Mythos 5. CNBC reported that Anthropic would retain all traffic on those models to defend against misuse and “complex and novel” cyberattacks. The company also pledged not to use the data for any “non-safety-related purpose,” model training included.

The help center page put the policy into effect on June 9. It covered zero data retention (ZDR) workspaces in Claude Console, Claude Code with ZDR in Claude Enterprise, and ZDR access through AWS Bedrock, Google Cloud Agent Platform or Microsoft Foundry. Free, Pro and Max plans were not affected, because those surfaces already retain inputs and outputs. Fable 5 and 5.1 run on the same underlying model as Mythos 5 and 5.1, with extra safeguards in the cyber and bio domains.

Anthropic’s reason for retention was that some attacks only show up across many requests. Best-of-N jailbreaking, for example, sends hundreds of slight variations of a single prompt. State-sponsored espionage and data extortion campaigns surface only when the safeguards classifiers can “zoom out across many requests.” Retained data was deleted after 30 days unless it was flagged or had to be kept for legal reasons. Humans could review it only through a controlled path, and every access went into a tamper-proof log.

In its post on the replacement, Anthropic calls it Enterprise Frontier Safeguards (EFS). EFS keeps ZDR-level privacy while still detecting misuse, because the data sits in cloud infrastructure the customer controls, not in Anthropic’s. Anthropic says it built EFS with more than 100 customers and with AWS, Google Cloud and Microsoft Azure. Participants included the Analysis and Resilience Center for Systemic Risk, whose members include the CISOs of the largest US banks, plus Comcast, KPMG, Mastercard, Salesforce and Visa. According to the company, enterprises generally understood why retention helped safety, but many of them, especially in regulated industries, found it hard to use models that kept their data.

On August 19, OpenAI said its business customers can keep zero retention and still get safety monitoring, with customer data allowed to stay on infrastructure the customer controls.

Until EFS is ready, eligible Anthropic customers get ZDR on Fable 5 and Fable 5.1.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.