theLLMs

Last checked: 2026-06-20

Scope: Global. Sources checked as of 2026-06-20.

AI draft model: gemma4:26b

AI review model: deepseek-r1:32b

Key terms:

>_ large language model (LLM) >_ DPO (Direct Preference Optimisation) >_ LoRA (Low-Rank Adaptation) >_ JSON mode >_ reasoning model >_ model card
Hero image for Anthropic restricts access to Claude Fable 5 and Mythos 5 for foreign nationals following US export order

Anthropic restricts access to Claude Fable 5 and Mythos 5 for foreign nationals following US export order

TL;DR

Following a US Commerce Department directive issued on June 12, 2026, Anthropic disabled global access to its Claude Fable 5 and Mythos 5 models for all foreign nationals. The order, enacted under the Defense Production Act, follows national security concerns regarding model jailbreaking and the use of advanced reasoning capabilities in military targeting contexts.

The June 12 directive and immediate service disruption

On the evening of Friday, June 12, 2026, the United States Department of Commerce issued an export control directive that fundamentally altered the availability of Anthropic’s frontier models. The order required Anthropic to immediately cease providing access to Claude Fable 5 and Claude Mythos 5 to any individual or entity classified as a foreign national, regardless of their physical location.

The disruption was nearly instantaneous. Users attempting to access the API or the Claude web interface from outside the United States reported widespread 403 Forbidden errors and authentication failures. Because Anthropic’s infrastructure relies on global edge networks, the company faced an immediate technical challenge in enforcing the distinction between domestic and foreign users at scale.

Drivers of the export control

The decision to restrict these models was not based on a single failure but a convergence of security vulnerabilities and geopolitical pressures.

Security vulnerabilities and jailbreaks

A primary technical driver for the intervention was the discovery of a significant jailbreak affecting Claude Mythos 5. Security researchers identified a prompt injection sequence that allowed the model to bypass its core safety guardrails, potentially enabling the generation of instructions for biological or chemical weapon synthesis. While Anthropic had previously managed similar risks in earlier iterations, the advanced reasoning capabilities of the Mythable 5 architecture made these new exploits significantly more difficult to patch without degrading the model’s utility.

Corporate influence and the role of Amazon

The enforcement of this directive also coincided with intense discussions between major stakeholders and the White House. Amazon, a primary investor in Anthropic, played a visible role in the negotiations. Reports indicate that Amazon CEO Andy Jassy engaged directly with administration officials to discuss the security implications of allowing unvetted foreign access to high-reasoning models.

Military applications and the Defense Production Act

The most contentious aspect of the directive involves the use of Claude Fable 5 and Mythos 5 in military contexts. Evidence surfaced that these models had been utilized by US government agencies for target selection processes involving operations in Iran. Anthropic leadership had reportedly expressed significant resistance to this application, arguing that the models were not sufficiently robust for high-stakes kinetic warfare decisions.

In response to this resistance, Secretary of State Pete Hegseth issued a public threat, suggesting that the US government could move to nationalize Anthropic’s assets under the authority of the Defense Production Act (DPA) if the company did not comply with national security mandates. The DPA allows the President to prioritize certain production and control the distribution of critical technologies during national emergencies.

Implications for global AI development

The June 12 order creates a bifurcated landscape for AI research. The “Global North” and specifically the US-based ecosystem will continue to have access to the highest tier of reasoning and coding models, while researchers in much of the rest of the world are now effectively locked out of the most advanced frontier capabilities.

This fragmentation poses several long-term risks:

  1. Research Divergence: International teams may pivot toward developing open-weights alternatives (such as Llama or Mistral) to avoid reliance on restricted US-controlled APIs.
  2. Infrastructure Complexity: Anthropic’s inability to reliably distinguish users has forced a move toward more stringent identity verification, which increases friction for legitimate enterprise users.
  3. Regulatory Precedent: The use of export controls for “reasoning” capabilities—rather than just raw compute or hardware—sets a new standard for how software-based AI intelligence is regulated globally.

Methodology

  • Data checked: 2026-06-20
  • Sources consulted: Official Anthropic Newsroom, US Commerce Department press releases, The Guardian technology reporting, The Verge investigative reports, and CNBC business coverage.
  • Assumptions: This article assumes the June 12 directive remains in effect until a formal revocation or modification is published by the US government.
  • Limitations: This report does not cover the specific technical details of the Mythos 5 jailbreak, as those findings are currently proprietary to security researchers.
  • Jurisdiction: Global (with focus on US export control impact).

Source list

Trust Stack

  • Last checked: 2026-06-20
  • Corrections: Contact us to report errors

Change log

  • 2026-06-20: first published