US orders Anthropic to disable AI models for all foreign nationals
A government order arrived at 5:21pm on a Friday. That timestamp matters. Bureaucracies love the hour when compliance teams want dinner and lawyers want ambiguity. Anthropic, maker of the Claude chatbot, says US agencies instructed it to suspend access to two advanced AI systems, Fable 5 and Mythos 5, for all foreign nationals. Not only abroad. Inside the United States too. Including foreign nationals working at Anthropic. The company says the letter invoked national security yet offered little concrete detail about the specific risk. That vagueness doesn’t just annoy. It forces guesswork and overreaction, because no firm wants to gamble against an export control directive.
A Switch Thrown in the Wrong Room
Anthropic frames the instruction as an export control directive. The phrase sounds neat until it hits daily life in tech. “Foreign national” covers visiting researchers, students, and long-term residents who still hold non-US citizenship. The order also reaches foreigners physically located in the US, which turns “export” into identity-based access control. Compliance becomes a sprint. Decide who counts. Block access fast. Prove it happened. Avoid locking out the wrong users. Anthropic says it had to cut off access on short notice. Regulators want immediate control. Companies respond with immediate disruption.
Mythos 5 and the Cyberweapon Fear
Mythos 5 sits at the center of the anxiety because it reportedly excels at finding software vulnerabilities, including old ones that survived for decades. Defenders love that. Attackers would love it more. The same capability that helps plug holes can also map them. US authorities and selected corporate partners have used the model to harden systems, which admits something important. The government already treats this AI as a strategic tool, not just a product. The dread comes from a familiar place. Cryptography and zero-day research triggered the same panic. Tools that shift power attract control. That impulse doesn’t wait for perfect evidence.
Fable 5, Safety Blocks, and a Thin Trigger
Fable 5 complicates the story because Anthropic says it blocks cybersecurity and biotechnology capabilities, despite drawing on Mythos technology. That sounds like an effort to ship usefulness while fencing off the sharpest edges. Anthropic also says it reviewed a report that likely triggered the directive, and its experts concluded the issue involved a limited ability to review program code and correct errors. That rationale feels thin because code review and bug fixing now sit inside many mainstream models. Anthropic even points at rivals, including OpenAI’s GPT-5.5, as sharing similar capabilities. If code review qualifies as the danger, the category expands until it captures most advanced assistants.
Nationality as a Security Policy
The directive does not say “malicious actors.” It says “foreign nationals.” That treats nationality as a shortcut for intent. Security people know shortcuts create errors in both directions. A serious attacker can route access through compromised accounts or intermediaries. A serious defender can hold the wrong passport. The order reportedly affects foreign employees at Anthropic, which turns a lab into a checkpoint and collaboration into a permissions problem. Earlier this month, Anthropic urged top AI companies to coordinate on a possible pause in advanced development, warning that capability growth could outrun human control. A pause proposal and an export restriction share a family resemblance. Both confess the pace feels dangerous.
Anthropic says it has only partial information from the government. That admission exposes the weakness. When regulators cite national security but skip specifics, they force firms into compliance theatre. Shut it down first. Ask what happened later. The public then learns a harsher lesson. Access to powerful AI can vanish overnight based on a letter and a definition of who counts as “foreign.” Meanwhile the technical reality keeps moving. Models that read code, find bugs, and propose fixes already spread across the market, and the line between “assistant” and “offensive tool” refuses to sit still. Nationality-based blocking may satisfy a legal template, yet it fails as a clean risk model because it confuses identity with behavior. Expect more sudden directives, and products designed for abrupt restriction.