Dozens of cybersecurity professionals have signed an open letter urging the White House to lift export control restrictions on Anthropic's frontier models, Fable and Mythos. Their argument: keeping the most capable defensive AI tools away from US allies makes everyone less safe. The government's argument: these models are too powerful to let out. Both sides are right, which is the problem.
The Borders of the Unthinkable
Export controls were designed for hardware: chips, warheads, dual-use components with physical mass. Applying them to language models introduces a category error that is producing visible friction. A 2026 arXiv paper on omnimodal agent orchestration describes AI systems operating across multiple modalities simultaneously, systems that are, by their nature, globally distributed and jurisdictionally ambiguous. You cannot export-control a distributed inference architecture the way you embargo a weapons shipment. The concept does not transfer cleanly.
When Knowledge Itself Is the Dual-Use Problem
What makes this story stranger is the parallel it draws with the history of cryptography. In the 1990s, the US government classified strong encryption as a munition. Mathematicians and civil libertarians fought back. The policy collapsed under its own absurdity. The Anthropic ban is the same fight, twenty years later, with higher stakes and less public attention. Big Tech's simultaneous lobbying blitz on AI regulation in Congress only muddies the water further: companies want deregulation at home while the government exports restriction abroad. Brewster Kahle spent decades arguing that access to knowledge is itself infrastructure. The Anthropic export ban is what happens when that infrastructure gets treated as a weapon.