New Security Concerns Emerge Over Anthropic AI Models

The United States government has introduced new export restrictions on Anthropic’s most advanced artificial intelligence models, known as Mythos 5 and Fable 5. These rules, which limit how the technology can be shared or sold outside the country, follow a period of growing concern regarding the safety and security of these powerful systems. While the company has already disabled access to these models in response to government orders, new details suggest that the decision was influenced by fears of foreign involvement.

According to a report from Semafor, officials in the White House were worried that a group linked to China might have gained unauthorized access to the Mythos model. If the Chinese government were able to use Mythos 5 or Fable 5, it could present a serious national security risk. Experts note that such access could allow a foreign power to copy the model’s behavior through a process called distillation. In this method, a smaller "student" AI is trained using the information from a more advanced model to copy its capabilities. By doing this, a country could potentially build its own version of the powerful technology without needing to develop it from scratch.

This is not the first time Anthropic has faced questions about the security of its most advanced software. The company has previously stated that models like Mythos are too powerful for the general public to use. However, reports have surfaced showing that a group on the messaging platform Discord was able to access the software for two weeks before the company discovered the breach and blocked them. These past incidents have made the government even more cautious about how such technology is protected.

There are different views on why these new restrictions were put in place. While the report from Semafor highlights the concerns about China, other voices have focused on different technical risks. For example, a post on the platform X by David Sacks, an advisor to Donald Trump, did not mention China. Instead, Sacks pointed to reports that the Fable and Mythos models could be "jailbroken." This term refers to finding ways to bypass the safety rules that AI companies put in place to prevent their models from being used for harmful activities. Anthropic has denied that its models can be easily broken in this way.

Anthropic has not provided a full response to these specific claims. However, a spokesperson for the company told Semafor that the government did not mention China during their discussions about the new export rules. This creates a complex situation where the exact reasons for the government's actions remain a subject of debate. The White House has also not officially confirmed the reports regarding the China-linked access.

As the situation develops, the focus remains on how companies like Anthropic can keep their most powerful tools safe. The ability of these models to process vast amounts of information makes them valuable, but also dangerous if they fall into the wrong hands. The government’s decision to limit access is a clear sign that officials are treating these AI models as a matter of national importance. For now, the models remain offline as the company works to meet the security requirements set by the government. The future of these advanced tools will likely depend on whether Anthropic can prove that it can keep its technology secure from both accidental leaks and intentional interference from foreign groups.