Particle.news
Download on the App Store

U.S. Order Forces Anthropic To Disable Its Two Most Advanced AI Models

U.S. officials say a possible jailbreak that could expose software vulnerabilities led to an export-control order that bars foreign nationals from using the models.

Overview

  • Anthropic said it disabled Fable 5 and Mythos 5 worldwide to comply with a Commerce Department export-control directive that requires blocking access by any foreign national.
  • The government cited a reported method to bypass the models’ safeguards that could let the systems find or reveal software vulnerabilities useful for cyberattacks, but it has not released detailed technical findings.
  • Reporting says Amazon researchers demonstrated ways to elicit vulnerability information from Fable 5 and that Amazon CEO Andy Jassy alerted U.S. officials, a development officials weighed when deciding to restrict access.
  • Anthropic criticized the process as opaque and disproportionate, said it received no formal technical disclosure, and is working to restore access while other Claude models remain available.
  • The action creates a new precedent for model-level export controls with immediate effects on Project Glasswing partners, foreign employees and researchers, cloud providers and companies planning public offerings.