Anthropic: Tensions Over AI Mythos with the US Government

Anthropic Facing Government Pressure

Last April, Anthropic unveiled an artificial intelligence model named Mythos, designed to interact with computer code at such an advanced level that it could potentially pose significant risks to global cybersecurity. To assess these risks, Anthropic allowed a select group of cybersecurity experts access to Mythos. Shortly thereafter, a modified version, dubbed Fable, was made public on June 9, presented as being safer. However, by the following Friday, the U.S. federal government deemed Fable a threat to national security, imposing restrictions on its export. In response, Anthropic quickly revoked access to both models.

The Influence of "Doomers" and Government Reaction

For several years, "doomers," those who worry about the catastrophic consequences of AI, have advocated for government intervention. Their wish came true, not due to an AI spiraling out of control or a biological weapon, but because of a particularly high-performing AI model in coding. This intervention seems more like a hasty reaction than a well-structured security plan.

The Implications of Amazon's Intervention

A notable element of this situation is the intervention of Andy Jassy, CEO of Amazon, who alerted authorities to the potential dangers of Fable. Amazon, having interests in Anthropic while also developing its own competing models, played a crucial role in this government decision. It is possible that this ban is temporary and may not withstand thorough legal scrutiny, especially since it is uncertain whether access to Fable can be considered an "export" in the legal sense.

International Consequences and China's Role

The ban has already had international repercussions. Some, like French politician Bruno Retailleau, see it as a "wake-up call" for Europe to strengthen its AI development. However, competition with China complicates this ambition. High-performing and cost-effective Chinese open-source models can be used without restrictions, attracting not only companies looking to avoid unilateral decisions from Washington but also cybercriminals. This situation could push American and European companies to turn to these Chinese models, as evidenced by the soaring stock prices of the Chinese startup Zhipu.

Cybersecurity Risks and the Non-Proliferation Debate

Paradoxically, restricting access to Anthropic's models could increase the vulnerability of the United States to cyberattacks. Cybersecurity experts have expressed their concerns in an open letter, stating that Anthropic's models help develop effective defenses. They believe these models are no more dangerous than other advanced technologies available on the market. This situation illustrates the challenges of applying the concept of non-proliferation to software, an approach typically reserved for nuclear weapons.

Political Reactions and the Future of Regulation

Finally, the reaction of American lawmakers is worth watching closely. The recent dispute between Anthropic and the government, particularly regarding the use of its models by the Pentagon, has already led to the introduction of new bills aimed at regulating military AI usage. Currently, tech companies and the White House dominate the AI debate. While discussions about increased federal regulation are ongoing, lawmakers struggle to define clear rules, particularly regarding the use of chatbots by children and the government's verification of AI model safety. Each decisive action from the White House intensifies the pressure for stricter regulation.

In summary, forecasts remain uncertain as the U.S. administration's positions on AI are constantly evolving. Under Trump, the focus was on deregulation to foster technological innovation. Today, the White House views the most valuable AI startup as a threat to national security. The coming months could bring new surprises in this ever-evolving field.