Particle.news
Download on the App Store

U.S. Limits Access to Leading AI Models as OpenAI Begins Restricted Preview

The administration says narrow jailbreak tests showed the models can reveal software flaws, so it is vetting users customer by customer while agencies and firms negotiate a permanent release process.

Overview

  • OpenAI began a U.S.-only, limited preview of its GPT-5.6 series after the company briefed the government and agreed to share the identities of approved partners for a staged rollout.
  • Anthropic earlier received a Commerce export-control order that forced a global shutdown of Fable 5 and Mythos 5 and has since been allowed to restore Mythos 5 for a vetted set of U.S. organisations while wider access remains restricted.
  • U.S. officials say the move responds to demonstrations that cutting-edge models can be 'jailbroken' to identify software vulnerabilities that hostile actors could exploit.
  • The interventions have prompted legal and diplomatic pushback, including a lawsuit challenging the export directive and talks with allied governments seeking trusted‑partner access.
  • Companies warn the ad-hoc gating disrupts cyber defenders and product rollouts, and both industry and agencies say they are working to turn the president's voluntary review order into a repeatable framework for future releases.