Why AI Export Controls Will Fail Just Like Encryption and Spyware Restrictions Did

Anthropic's Mythos Ban: The Opening Shot in America's AI Export Control Battle

The U.S. government's attempt to clamp down on AI export controls has arrived — and it has landed with all the subtlety of a sledgehammer. The White House recently ordered Anthropic to halt exports of its powerful AI models, Fable and Mythos, to anyone outside the United States, including foreign nationals already inside the country, citing unspecified national security concerns. Anthropic was reportedly given approximately 90 minutes to comply before having to pull both models from availability entirely. Neither model has been accessible to anyone since.

For developers, IT decision-makers, and policy professionals tracking the intersection of AI capability and government regulation, this moment carries enormous weight. It is the first serious test of whether Washington can use export control law — a legal mechanism born in an era of physical goods and weapons — to contain what is arguably the most complex, intangible, and rapidly proliferating technology in human history. As TechCrunch reports, the outcome of this standoff will not just affect Anthropic's market access — it will define the compliance framework every major AI lab must navigate going forward.

What Actually Triggered the Mythos Restriction?

The immediate catalyst appears to be two separate incidents that alarmed U.S. officials in quick succession. First, Anthropic granted a South Korean telecommunications company access to Mythos through its limited partner program. U.S. intelligence officials reportedly flagged this company — widely identified in press reports as SK Telecom, which has denied any China connection — as having suspected ties to Beijing. The mere possibility of an adversarial state gaining access to a frontier AI model purpose-built for cybersecurity offense and defense was apparently enough to trigger alarm at the highest levels.

Second, Amazon CEO Andy Jassy reportedly alerted the administration after Amazon's own researchers claimed to have found a method to circumvent the safety guardrails built into Fable 5. Anthropic pushed back firmly on this characterization, describing the issue as a narrow, already-patched vulnerability rather than a wholesale compromise of the model's safety architecture. Regardless of who was right on the technical merits, the political result was the same: the Commerce Department issued an export control directive, and Anthropic found itself scrambling to shut down access to both models within the time it takes to run a brief stand-up meeting.

Mythos had never been a wide-release product to begin with. Since its April launch, Anthropic had restricted access to approximately 150 vetted companies and government organizations — the premise being that a model with such potent cybersecurity capabilities should only be accessible to defenders trying to harden systems before malicious actors developed equivalent capabilities on their own. That controlled-access philosophy clearly did not satisfy regulators once geopolitical risk entered the picture.

The Crypto Wars Taught Us That Restricting Software Is Fundamentally Different From Restricting Hardware

To understand why this effort is likely to falter, you need to go back to the early 1990s, when the U.S. government convinced itself that encryption was a weapon it could control at the border. Phil Zimmermann had developed Pretty Good Privacy (PGP), a software tool that allowed ordinary internet users to encrypt their communications so effectively that even intelligence agencies could not intercept and read them. The U.S. Customs Service opened a criminal investigation against Zimmermann for allegedly violating arms export controls — treating a software program distributed over the nascent internet as if it were a shipment of missiles.

Zimmermann's response was as clever as it was principled: he published PGP's source code as a printed book. Under the First Amendment, books could be exported. The government's attempt to contain encryption collapsed under the weight of its own logical inconsistency. The investigation was eventually dropped, and the episode — now known as the "Crypto Wars" — paved the way for the end-to-end encryption that now protects billions of users on platforms like Signal and WhatsApp. As the Electronic Frontier Foundation has long documented, the lesson was clear: software, once created, does not respect borders.

How the Wassenaar Arrangement Failed to Contain Spyware — and Why AI Faces the Same Pitfalls

The next major experiment came in the early 2010s, when researchers began uncovering Western-made surveillance software being deployed against journalists and dissidents across the Middle East. In response, participating nations agreed to expand the Wassenaar Arrangement — an international treaty that governs the export of dual-use technologies — to include surveillance and intrusion software. Companies would now need export licenses to sell their hacking tools abroad.

The initiative had two structural weaknesses that proved fatal. First, major spyware-producing nations like Israel never signed on to the arrangement, creating immediate jurisdictional blind spots around some of the world's most capable and commercially active surveillance vendors. Second, even among signatory nations, enforcement was left entirely to the discretion of national governments. Italy, for instance, granted export licenses to Hacking Team — a company whose tools were documented to have been used by authoritarian regimes against civil society — despite that track record being publicly known at the time.

Europe's record has remained troubling. Despite repeated scandals and considerable public attention, the European bloc has consistently struggled to prevent spyware exports to authoritarian states. Critics of the EU's most recent initiative in this space, as Reuters has reported, argue it "does not go far enough." Companies like the sanctioned Intellexa consortium simply restructured and relocated to jurisdictions with weaker controls. Others explored moving operations to Saudi Arabia for similar regulatory arbitrage reasons. The one clear success story — German spyware maker FinFisher shutting down in 2022 following a multi-year prosecution for allegedly selling tools to Turkey without a license — is notable precisely because it was exceptional rather than routine.

Why AI Export Controls Face Even Steeper Obstacles Than Their Predecessors

If encryption and spyware proved difficult to contain through export controls, AI models present an even more complex challenge. Consider the core differences in what regulators are actually trying to restrict.

Unlike a physical export, AI capabilities can travel through API calls, model weights, and fine-tuning techniques. Researchers in Europe, China, and elsewhere are developing their own frontier models independently. The open-source AI ecosystem — including powerful open-weight models that can be downloaded and run locally — means that restricting one commercial product does not close the capability gap. It simply inconveniences the legitimate, compliant users while bad actors route around the restriction entirely.

There is also the competitive cost to consider. According to analysis published by the Brookings Institution on AI export controls, overly broad restrictions risk ceding market share to international competitors without meaningfully delaying adversarial AI development. The U.S. is not the only country building powerful AI. For European businesses and government organizations currently evaluating AI tools and their compliance posture under GDPR and digital sovereignty frameworks, this instability in the U.S. AI market is itself a data point worth factoring into procurement decisions.

What This Means for Developers, Privacy Professionals, and Organizations Using AI Tools

The immediate practical consequence of the Anthropic restrictions was experienced by the roughly 150 partner organizations that lost access to Mythos without meaningful notice. That 90-minute compliance window is a sobering reminder that dependency on a single vendor's frontier AI model — particularly one operating under a restricted-access program with government visibility — carries tail risks that most operational risk frameworks do not currently account for.

For organizations operating under GDPR, digital sovereignty mandates, or EU data residency requirements, the episode adds another dimension to the argument for diversifying AI tool dependencies. European-built or open-source AI alternatives may lack the raw capability of frontier U.S. models today, but they offer something increasingly valuable: regulatory stability and jurisdictional clarity. A model that cannot be switched off by a foreign government's executive order is, in certain operational contexts, more reliable than a more capable one that can.

The broader policy trajectory also matters. If the Trump administration maintains the restrictions, U.S. AI companies may require government pre-approval before onboarding foreign customers — a compliance burden that would significantly affect pricing, speed-to-market, and the attractiveness of American AI products in European and Asian enterprise markets. If the administration reverses course, as some analysts consider likely given the competitive implications, it will represent a tacit acknowledgment that frontier AI capabilities will proliferate regardless of what Washington restricts. Wired's coverage of related chip export controls has previously illustrated how even hardware restrictions at the semiconductor level have proven difficult to enforce as a long-term containment strategy.

What Three Decades of Failed Cyber Export Controls Tell Us About the Road Ahead

The pattern that emerges across the history of encryption regulation, spyware treaties, and now AI export controls is consistent: governments tend to overestimate their ability to contain software-based capabilities and underestimate the speed at which the global research community independently develops equivalent tools.

Originally reported by TechCrunch. Summarised and curated by European Purpose.