Why US Export Controls on Mythos Are Likely to Fail Like Encryption and Spyware Before It

The White House ordered Anthropic to restrict the export of its AI models Fable and Mythos to anyone outside the United States, as well as to foreign nationals inside the country, citing unspecified national security concerns. After the order, Anthropic pulled both models from availability, rendering them inaccessible for a week. The episode represents the first major test of whether export controls can contain frontier AI in ways previous attempts to restrict encryption and spyware failed to do—and the outcome may determine how governments regulate AI capabilities for decades to come.

What Triggered the Mythos Ban

Anthropic launched Mythos in April, marketing it as a cyber tool that could identify and remediate vulnerabilities before adversaries exploit them. The company restricted access severely: only around 150 vetted companies and government organizations had access to the model prior to the ban.

Two events reportedly triggered the export restriction. First, Anthropic gave a South Korean telecom access to Mythos through its limited partner program, and U.S. officials grew alarmed after identifying the company as one they suspected had ties to China—though the company, widely reported to be SK Telecom, has denied any China connection. Second, Amazon CEO Andy Jassy reportedly alerted the administration after Amazon's researchers found a way around Fable 5's safeguards. Anthropic disputes the "jailbreak" label, calling it a narrow, already-patched issue rather than a wholesale defeat of the model's safety measures.

The Commerce Department issued an export-control directive, and Anthropic had to scramble to immediately limit access to its products—within roughly 90 minutes of being notified, by some accounts.

The Failed Pattern: From Encryption to Spyware

Governments have tried to use export controls to limit the proliferation of what they see as dangerous cyber technology for decades, but their track record has been middling at best. The history offers stark lessons for why Mythos restrictions are unlikely to succeed.

The first major failure came during encryption wars in the early to mid-1990s. Computer scientists were developing encryption technologies to secure data as it traveled over the internet, including Pretty Good Privacy (PGP), a popular software that could encrypt data and make it virtually impossible to unscramble even if intercepted. The U.S. government initially saw PGP as a dangerous weapon, fearing it would prevent its intelligence agencies from snooping on emails, and the U.S. Customs Service opened a criminal investigation against PGP's creator Phil Zimmermann for allegedly violating arms export controls.

Zimmermann fought back by publishing PGP's source code as a printed book, igniting what is known today as the "Crypto Wars." Zimmermann later won when the investigation was closed, paving the way for crucial end-to-end encryption algorithms such as the one used by billions of Signal and WhatsApp users. The control regime collapsed because the technology was impossible to keep proprietary once published, and enforcement became practically impossible.

The second failure involved commercial spyware. During the early 2010s, researchers began discovering Western-made spyware used against dissidents in the Middle East. In response, several governments agreed to expand the Wassenaar Arrangement, an international treaty that limits the export of dual-use software and technologies used in both civilian and military applications, classifying surveillance and hacking software as dual-use and forcing spyware makers to get export licenses to sell their products abroad.

But the Wassenaar framework proved ineffective for multiple reasons. Wassenaar has always had two inherent weaknesses: several countries that don't adhere to the agreement, including Israel, which houses some of the world's most active spyware makers, and the agreement depends on countries applying it to companies within their borders at their own discretion. For a time, the Italian government allowed one of the country's then-top spyware makers, Hacking Team, a license to export its tools around the world, despite the company's track record of selling spyware to oppressive governments that used it to hack journalists and human rights activists.

Despite numerous scandals, Europe, home to many spyware makers, has continually failed to curb the export of spyware to authoritarian regimes, with critics saying that a recently renewed effort across the bloc of 27 member states does not go far enough. Several spyware makers, such as Intellexa, a sanctioned consortium of spyware companies, have simply moved their operations to countries with lax export controls, while other spyware makers sought to move their operations to Saudi Arabia for similar reasons.

One exception proved limited: Germany-based spyware maker FinFisher shut down in 2022 after a multi-year investigation by German prosecutors into the company for allegedly selling spyware to Turkey without an export license, with investigators previously finding the FinFisher spyware had been deployed on the phones of critics of Turkey's government. Yet this single success against one company contrasts sharply with the persistent proliferation of spyware globally despite decades of regulatory effort.

Why AI Software Controls Face the Same Structural Barriers

The structural problem underlying both failures applies with even greater force to Mythos. Software and AI model weights lack the physical traceability that makes hardware export controls partially enforceable. A fabrication tool like an ASML extreme ultraviolet lithography machine is a single, traceable, heavily controlled object. By contrast, an AI model capable of generating offensive cyber capabilities can be compressed, transferred, and reproduced without the physical chokepoints that make chip controls even partially effective.

The US approach to AI export controls has focused heavily on restricting advanced semiconductors. The Commerce Department's restrictions on advanced GPUs—targeting Nvidia's H100 and subsequent architectures—represent the most aggressive use of export controls in the semiconductor sector in decades. This hardware-focused strategy has real leverage because semiconductors are physical goods that require specialized manufacturing infrastructure. According to MachineHours, TSMC controls 92% of production for the most advanced chips at 7nm and below, the chips that power the AI systems at the center of the export control debate. This concentration gives the US meaningful control over hardware supply chains, which is why GPU export restrictions carry real weight.

Software controls operate under fundamentally different enforcement logic. Once an AI model's weights exist and researchers outside US jurisdiction have access to analogous training infrastructure, reproduction and distribution become difficult to prevent through regulatory channels alone. The technology can be copied, distributed through private channels, and replicated by others with access to sufficient computational resources. The ease of replicating AI model weights and the global availability of the underlying compute and training data mean restrictions on any single tool or company are unlikely to contain the capability itself.

As of the time of writing, the impasse between Anthropic and the Trump administration remains unresolved. There is a reasonable chance the administration will buckle and lift the restriction in the interest of keeping American AI companies competitive worldwide—a move that would amount to tacit acknowledgment that AI labs elsewhere, including in China, will likely reach similar capabilities regardless of what the U.S. restricts. Alternatively, American AI companies could end up needing government approval before serving foreign customers at all, a compliance burden that would invariably dent their bottom line.

Strategic Implications for US Technology Policy

The distinction between hardware and software controls matters enormously for US strategy. Washington can slow China's access to leading-edge chips by pressuring ASML and TSMC through their export dependencies. It cannot apply the same logic to AI software once a capability exists and researchers outside US jurisdiction have access to analogous training infrastructure. Understanding this distinction is essential for coherent technology policy going forward.

Given the past experiences that world governments have had with trying to control the reach of software, government-mandated export controls are unlikely to be the right approach to stop malicious actors from abusing powerful dual-use cyber technologies. The historical record on encryption and spyware export controls suggests that the Mythos ban represents the continuation of a pattern rather than a departure from it: regulatory attempts that are "simultaneously over-broad and under-effective," constraining American companies while failing to meaningfully contain the underlying capabilities from reaching adversaries.