Source linked

Claude Fable 5 en Bedrock: datos compartidos con Anthropic, luego revocados

Anthropic obligó a los usuarios de Amazon Bedrock a optar por compartir prompts y outputs para la revisión humana, luego pidió a AWS que revocara el acceso a ambos Claude Fable 5 y Mythos 5 tres días después del lanzamiento citando el control de exportación de Estados Unidos.

anthropicamazon bedrockclaude fable 5mythos 5ai data privacyexport controls

Three days after Amazon Bedrock launched Claude Fable 5 and Mythos 5, Anthropic asked AWS to revoke access to both models, citing US export control compliance.

Data Sharing Was a Non-Negotiable Requirement

Using either model on Bedrock required opting into provider_data_share. That setting sends every prompt and output to Anthropic, where they retain the data for 30 days and subject it to human review. This is a sharp departure from previous Bedrock model deployments, where inference data stayed inside the AWS boundary and never crossed to the model provider.

Anthropic made no promises about opt-out paths. If you wanted to run Fable 5 or Mythos 5 on Bedrock, you had to accept that your organization's traffic would be read by Anthropic personnel. No other configuration was available.

The Surprise Revocation Three Days Later

On June 20, 2026, Steef-Jan Wiggers reported that Anthropic had asked AWS to remove both models from the Bedrock marketplace, citing US export control compliance. The exact regulations at issue were not specified, but the timing is telling: a launch, then a pullback within 72 hours. This suggests either Anthropic belatedly realized some legal exposure or AWS flagged a compliance gap after initial deployment.

Users who had already integrated the models now face a sudden loss of access. No migration path to a compliant version has been announced. Anthropic's own API may still offer these models under different data handling terms, but the Bedrock integration is effectively dead.

What This Means for Bedrock Users

If you are evaluating models on Bedrock, the Fable 5 episode is a case study in the risks of relying on cloud marketplace model deployments. Data-sharing terms can change, and models can vanish without notice. Always read the fine print on provider_data_share and have a fallback plan that does not depend on a single provider's API endpoint. Expect future model integrations to include explicit data flow audits and contractual guarantees about retention windows, because this won't be the last time legal compliance overrides product availability.


Source: Claude Fable 5 on Bedrock Requires Sharing Inference Data with Anthropic
Domain: infoq.com

Read original source ->

External source stays available while the OJO article and comment thread stay local.

Comments load interactively on the live page.