Anthropic reverses course: keep your Claude logs in your own cloud

The retention Anthropic forced on you for safety is gone after pushback — but you now run the misuse-monitoring it used to. Also: OpenAI–Epic, Mistral OCR.

Nowline Sep 2 1:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The forced-retention era just ended

    Anthropic dropped the temporary data retention it required for safety checks — the policy The Register says cost it enterprise deals — after customer pushback. Its replacement, Enterprise Frontier Safeguards (EFS), landed Sept 1. If Claude holding onto your prompts was a blocker, that objection is now addressable.

  • Your logs live in your bucket, with your keys

    Activity data goes to your own AWS S3, Azure Blob, or Google Cloud Storage — you hold the encryption keys and access policies, and no Anthropic employee reviews it; misuse alerts route straight to your security team. It's free from Anthropic; you just pay your cloud's storage and egress.

  • The catch: the monitoring is now your job

    Zero-data-retention isn't automatic — you apply for it, and once EFS is on, its cross-session misuse detection (credential theft, cyberattack patterns) sends signals to you to review, not to Anthropic. The Register's blunt read: you now have to check it actually worked.

  • What's covered, and when

    EFS spans Claude Code, Claude Platform and Enterprise, plus Amazon Bedrock, Google's Agent Platform, and Microsoft Foundry, on Fable 5, Fable 5.1, and Mythos-class models. Rollout is phased through fall 2026; eligible customers get interim ZDR on Fable 5/5.1 until it ships.

  • Elsewhere: OpenAI pipes Epic records into ChatGPT

    ChatGPT Health can now import a patient's Epic EHR data (read-only) and pull public sources like PubMed, OpenAI said Sept 1. It's clinician-facing, not a builder API — but it signals where the assistant-plus-your-own-data pattern is heading.

  • Elsewhere: Mistral OCR 4.1 hits GA

    mistral-ocr-4-1 is now production-GA, with a confidence_scores_granularity setting for page-, block-, or word-level scores — handy for flagging low-confidence extractions before they poison a RAG index.