Anthropic Mythos: Washington Fears Chinese Access

Washington Restricts Access to Mythos

According to information from Semafor, the recent export restrictions imposed by Washington on Mythos, the most powerful AI model from Anthropic, are partly motivated by fears that China may have gained access to it. This situation raises significant national security concerns. The U.S. government suspects that China could have used Mythos to train its own AI models, leading to these restrictive measures.

Anthropic considers Mythos too advanced and sensitive to be made available to the general public. It is reserved for a select group of trusted organizations. However, Washington's concerns have recently led to the imposition of export restrictions on this system. The true motivations behind this decision are beginning to emerge, and if suspicions of Chinese access are confirmed, it would far exceed a mere cybersecurity incident.

Suspicions of Chinese Access

According to a report from Semafor, the White House's decision to impose export restrictions on Mythos was fueled by fears that a group linked to China had accessed the model. If the Chinese government had indeed gotten hold of Mythos 5 or Fable 5, the implications for national security would be significant.

One concrete threat would be the use of distillation, a technique by which an AI model known as a "student" is trained from a more advanced system to replicate its behaviors. This process would theoretically allow for the replication of a cutting-edge model's capabilities without having developed the technology oneself.

The White House has not confirmed this information. In a post on X, David Sacks, an advisor to Donald Trump on AI issues, did not mention China, instead highlighting vulnerabilities related to jailbreaks, which are methods to circumvent the model's safeguards. Anthropic has denied these accusations, and a spokesperson for the company stated that the issue of China was not raised during discussions with the government regarding export restrictions.

Anthropic and Security Incidents

If Chinese access were to be confirmed, it would not be the first time that Mythos found itself at the center of an embarrassing breach for Anthropic. The company had previously faced a notable security breach: a Discord group gained unauthorized access to the model for two whole weeks before Anthropic detected the intrusion and severed the connection.

These successive incidents raise a fundamental question: how can a technology deemed too dangerous for the public end up, on two occasions, in unauthorized hands? For Anthropic, the answer to this question is crucial for both its credibility and its relationship with U.S. authorities.