A recent directive from the White House to US-based AI developer Anthropic, ordering the restriction of its advanced AI models, Fable and Mythos, based on undisclosed national security concerns, has reignited a long-standing debate. The company swiftly responded by making both models inaccessible, prompting discussions on whether such export controls can effectively manage the spread of powerful artificial intelligence, particularly given the chequered history of similar attempts with encryption and spyware.
Anthropic had previously marketed Mythos as a highly potent cybersecurity tool, capable of both defending against and potentially creating sophisticated cyber threats. Before the ban, access was limited to approximately 150 vetted organisations, with the intention of helping cybersecurity defenders pre-empt future threats. Reports suggest the ban was triggered by two key events: Anthropic providing a South Korean telecom company, suspected of ties to China, access to Mythos, and alleged vulnerabilities in Fable 5 discovered by Amazon researchers, which Anthropic disputes as a minor, patched issue.
This is not the first time governments have attempted to regulate the flow of sensitive digital technologies. In the early 1990s, the US government famously tried to classify encryption software, such as Pretty Good Privacy (PGP), as a munition. This led to the 'Crypto Wars', where PGP's creator, Phil Zimmermann, published its source code as a book to circumvent export restrictions. Ultimately, these efforts failed, paving the way for the widespread adoption of end-to-end encryption used in services like WhatsApp and Signal today.
More recently, in the early 2010s, the discovery of Western-made spyware being used against dissidents led to attempts to expand the Wassenaar Arrangement, an international treaty governing dual-use technologies. The aim was to classify surveillance software as dual-use, requiring export licenses. However, despite these measures, the proliferation of such tools has continued, raising questions about the practical enforceability and effectiveness of such controls in a globally interconnected digital landscape.
The current situation with Anthropic's models presents a crucial test for how governments will approach the regulation of frontier AI. The outcome could significantly influence not only Anthropic's access to international markets but also establish precedents for other AI development labs globally. The challenge lies in balancing national security interests with the potential for technological advancement and the inherent difficulty of containing digital information in an age of rapid global dissemination.
For the UK, these developments carry significant implications. As a leading hub for AI research and development, the UK's AI sector could be affected by evolving international export control frameworks. Companies reliant on advanced AI models for cybersecurity, research, or commercial applications may face new restrictions or uncertainties. Furthermore, the broader debate on AI governance and responsible deployment will undoubtedly influence UK policy and regulatory approaches, including those from the Information Commissioner's Office (ICO) and the potential impact of the EU AI Act.
Source: White House, Anthropic