Key facts
- The White House has ordered Anthropic to restrict the export of its AI models Fable and Mythos due to unspecified national security concerns.
- Anthropic has made both models unavailable for approximately a week following the directive.
- The U.S. government's action is seen as a test of its ability to control the proliferation of advanced AI technologies through export controls.
- Historical attempts to control encryption and spyware exports have yielded mixed results.
- The decision could impact Anthropic's access to foreign markets and set a precedent for other AI developers.
The White House has directed Anthropic to halt exports of its advanced AI models, Fable and Mythos, citing national security concerns. This action, taken last Friday, has led to the models becoming unavailable for about a week and represents a significant test for the U.S. government's strategy of using export controls to manage the proliferation of powerful AI technologies.
Anthropic had marketed Mythos as a potentially dangerous tool capable of causing significant harm if widely released, with access previously limited to around 150 vetted companies and government organizations. The ban was reportedly triggered by two events: Anthropic granting access to Mythos to a South Korean telecom suspected of having ties to China, and Amazon CEO Andy Jassy alerting the administration after researchers allegedly found a way to bypass safeguards on Fable 5. Anthropic has disputed the characterization of the Fable issue as a 'jailbreak,' describing it as a narrow, patched vulnerability.
The Commerce Department issued the export control directive, requiring Anthropic to limit access to its products within approximately 90 minutes. This situation echoes historical U.S. government attempts to control the spread of sensitive cyber technologies, such as encryption and spyware, which have historically met with limited success.
In the 1990s, the U.S. government's efforts to restrict the distribution of encryption software like Pretty Good Privacy (PGP) ultimately failed when its creator, Phil Zimmermann, published the source code, leading to widespread adoption of end-to-end encryption. More recently, the Wassenaar Arrangement, an international treaty aimed at controlling dual-use technologies, has struggled to curb the export of spyware, with countries like Israel and Italy having lax enforcement, and companies relocating to jurisdictions with weaker controls.
The current impasse between Anthropic and the U.S. administration remains unresolved. There is a possibility the restrictions could be lifted to maintain the competitiveness of U.S. AI firms globally, acknowledging that similar capabilities may emerge elsewhere. Alternatively, U.S. AI companies might face increased compliance burdens requiring government approval for foreign customers, potentially impacting their profitability. The article suggests that government-mandated export controls may not be an effective long-term solution for preventing malicious actors from abusing powerful dual-use cyber technologies.