The US government's sudden ban on Anthropic's Fable 5 and Mythos 5 models reveals the actual weakness of export controls in the AI era: they can only work if applied uniformly across all builders, and that consensus doesn't exist. The jailbreaks Amazon researchers found in Fable 5 already exist in other models sitting on the market. Anthropic itself made this point public. Cybersecurity researchers signed an open letter opposing the ban. Three decades of failed attempts to restrict encryption and security software via export controls should have made this predictable. What the ban appears to accomplish instead is free marketing for Anthropic's brand and a visible reminder that regulatory action against one company while others ship identical capabilities looks arbitrary rather than protective.
Meanwhile, the real consolidation of power is happening through standardization, not restriction. Google, Microsoft, OpenAI, and others have created the Appia Foundation to establish specifications for AI behavior verification. Separately, Google, Microsoft, Cisco, Nvidia, and Salesforce are backing Agentic Resource Discovery, a protocol to standardize how agents access tools within corporate systems. These aren't restrictions. They're infrastructure plays that lock in the large players as the gatekeepers of how enterprise AI gets built and audited. The federal government is adopting these patterns too: the National Endowment for the Arts deployed an AI coding agent to rebuild its grants system in a week. NASA runs critical systems on the core Flight System framework. Federal agencies are moving quickly to adopt AI agents for mission-critical processes. When government and enterprise both standardize on frameworks controlled by the same five or six companies, export bans become almost beside the point.
The real tension is between builders who ship products and committees that publish frameworks. Noam Shazeer, a co-author of Attention Is All You Need, just moved from Google to OpenAI. Subquadratic emerged from stealth claiming to have solved a mathematical bottleneck holding back LLMs for a decade. Temporal is positioning itself as the reliability backbone for agentic AI by stabilizing failed processes through durable execution. These are companies shipping solutions to concrete problems. The Appia Foundation and ARD protocols are attempts to standardize the space before it fragments. One approach generates velocity and optionality. The other generates compliance and predictability. The ban on Anthropic's models achieved neither.
Sloane Duvall