US Issues Verbal Export Directive Blocking Anthropic Cybersecurity Models

The directive, issued verbally and citing an unverified jailbreak method, forced immediate withdrawal of both models despite Anthropic's assessment that the vulnerabilities were minor, previously known, and reproducible in GPT-5.5. No written order or CVE has been released, marking the first known use of export authorities against an AI model rather than hardware. Anthropic disputes the scope while complying, noting the standard would halt frontier deployments industry-wide.

The action arrived two days after CEO Dario Amodei called for statutory powers to block unsafe AI releases and weeks after Defense Secretary Pete Hegseth designated Anthropic a supply chain risk following failed military contract talks. This sequence reveals operational friction between voluntary safety advocacy and actual enforcement mechanisms, where verbal directives bypass transparency requirements applied to chip export controls.

Procurement records and prior BIS rules on advanced semiconductors show the government already tracks dual-use AI capabilities through compute thresholds; extending that logic to model weights without public criteria creates precedent for selective recalls. Independent technical confirmation of the claimed jailbreak remains absent, leaving the evidence trail limited to government assertion.

Restoration depends on whether a formal statutory process replaces the current verbal channel. Industry filings ahead of Anthropic's anticipated IPO will likely disclose ongoing compliance costs and model access restrictions.