Particle.news

Anthropic Locks Down Unpublished Files After CMS Error Exposes Drafts and Model Plans

The company says a CMS error briefly exposed only draft materials.

Overview

  • Fortune reported that Anthropic’s unpublished cache was reachable through an external content system, and the company restricted access after it was notified.
  • A University of Cambridge researcher counted close to 3,000 unpublished assets that were public by default unless set to private, and Anthropic blamed human error in the CMS configuration.
  • The exposed files described an unreleased AI model, and Anthropic confirmed it is testing a more capable system with early access customers that it characterized as a step change in reasoning, coding, and cybersecurity.
  • Other items included details of an invite-only CEO retreat in the U.K. and internal-use images, while Anthropic said no core infrastructure, AI systems, or customer data were involved and the issue was unrelated to its Claude tools.
  • India Today reported that a draft named the model “Claude Mythos” and flagged cybersecurity risks, which Anthropic has not confirmed, highlighting how misconfigured web tools and modern crawlers can surface unintended public assets.