Why U.S. Cyber Export Controls Keep Failing — And What Anthropic's Mythos Ban Tells Us

The U.S. Government Just Banned Anthropic's Most Powerful AI Models — Here's Why History Says It Won't Stick

For more than 30 years, governments around the world have attempted to bottle up dangerous cybersecurity software using export controls. The results have been, at best, mixed. Now the United States is applying that same playbook to artificial intelligence — specifically to Anthropic's advanced cybersecurity model, Mythos — and history offers little reason for optimism.

Last Friday, the White House directed Anthropic to immediately halt exports of two of its flagship AI models, Fable and Mythos, citing vague national security concerns. The directive extended to foreign nationals residing inside the U.S. as well. Within roughly 90 minutes of being notified, Anthropic pulled both models offline, leaving them inaccessible to everyone for over a week.

This marks the first genuine test of whether Washington can leverage export restrictions to rein in frontier AI — a question whose answer will not only determine Anthropic's future in global markets but also set a precedent that every other major AI lab will need to factor into its business strategy.

What Made Mythos So Controversial in the First Place

When Anthropic launched Mythos in April, it was never a product for the general public. The company positioned it as an extraordinarily powerful cybersecurity tool — one capable of identifying and exploiting software vulnerabilities at a scale that made even its creators cautious. Before any government ban, access was already limited to roughly 150 carefully vetted companies and government agencies. The idea was to put the technology in the hands of defenders so they could shore up their systems before malicious actors could develop equivalent capabilities independently.

So what pushed federal officials over the edge? Two events, reportedly. First, Anthropic extended Mythos access to a South Korean telecommunications company through its limited partner program. U.S. officials became alarmed when they identified the firm — widely reported to be SK Telecom — as a company they believed had ties to China. SK Telecom has flatly denied any such connection.

The second trigger came from an unlikely source: Amazon CEO Andy Jassy reportedly alerted the administration after Amazon's own researchers claimed they had found a way to bypass the safety guardrails of Fable 5, another Anthropic model. Anthropic pushed back on that characterization, describing the issue as a narrow, already-patched edge case rather than a fundamental security failure.

Regardless of who was right, the Commerce Department issued an export control directive, and Anthropic was forced to act almost immediately.

A Decades-Long Track Record of Failure

The uncomfortable truth is that export controls on cybersecurity technology have a poor track record. The most glaring example dates back to the early 1990s, when the U.S. government attempted to suppress a piece of encryption software called Pretty Good Privacy, or PGP.

Developed by programmer Phil Zimmermann, PGP allowed users to encrypt their digital communications so effectively that even intercepted messages were virtually impossible to decode. Federal authorities worried the tool would blind intelligence agencies trying to monitor communications and responded by opening a criminal investigation against Zimmermann under arms export control laws.

Zimmermann's countermove was both clever and defiant: he published PGP's source code as a physical printed book, making it legally protected speech under the First Amendment. The move ignited the so-called "Crypto Wars," a prolonged standoff between technologists and governments over the right to use strong encryption. Zimmermann ultimately prevailed — the investigation was dropped — and the encryption algorithms that emerged from that era now secure the private messages of billions of people on platforms like Signal and WhatsApp.

Spyware and the Limits of the Wassenaar Arrangement

The next major chapter in cyber export control came in the early 2010s, when security researchers began uncovering surveillance software built by Western companies deployed against journalists, dissidents, and activists across the Middle East and beyond. Governments responded by expanding the Wassenaar Arrangement, an existing international treaty designed to regulate dual-use technologies — those with both civilian and military applications.

The goal was to require spyware makers to obtain export licenses before selling their products to foreign governments. In theory, this would create a paper trail and a gatekeeping mechanism. In practice, the arrangement was riddled with structural weaknesses.

Two Core Problems That Undermined the Approach

First, not every country signed on. Israel, home to some of the world's most prolific spyware developers, has never been a party to the Wassenaar Arrangement. Second, even among signatory nations, enforcement was left entirely to each government's discretion — and many were remarkably permissive.

Italy, for instance, allowed its homegrown spyware firm Hacking Team to export its surveillance tools globally for years, even as evidence mounted that the company's clients included authoritarian governments using the software to target human rights defenders and members of the free press. Europe more broadly has been repeatedly criticized for allowing spyware companies to operate with minimal accountability. Critics of the European Union's most recent efforts to address the problem say the new measures still "do not go far enough."

Some companies simply relocated to countries with looser regulations when heat from authorities became too intense. Intellexa, a sanctioned spyware consortium, restructured its operations across multiple jurisdictions to stay a step ahead of enforcement. Others reportedly explored moving to Saudi Arabia for similar reasons.

There have been isolated victories. German prosecutors successfully shut down spyware company FinFisher in 2022 after a years-long investigation into the firm's alleged sale of surveillance tools to Turkey without a proper export license — tools that had been found on the phones of Turkish government critics. But such outcomes are the exception, not the rule.

What Happens Next with Anthropic and Mythos

As of now, the standoff between Anthropic and the Trump administration remains unresolved. Two plausible outcomes loom. The administration could walk back the restriction, effectively acknowledging that AI capabilities similar to Mythos will inevitably emerge in other countries — including China — regardless of what American regulators do. Alternatively, the government could formalize a system requiring AI companies to seek approval before serving international customers, a compliance burden that would add cost, complexity, and competitive disadvantage for U.S. firms.

Neither path is clean. And given what the past three decades have demonstrated about the limits of software export controls — whether applied to encryption tools or commercial spyware — there is little evidence that government-mandated restrictions alone can reliably prevent sophisticated technology from reaching the wrong hands.

The Mythos ban may ultimately be remembered less as an effective security measure and more as the moment the AI industry confronted a question the cybersecurity world has been wrestling with for a generation: can any government truly contain powerful dual-use technology once it exists?