Anthropic pulls two new AI models after U.S. regulators raise concerns

Anthropic said it is immediately suspending access to two of its most advanced AI models, just days after opening them to the public, after U.S. authorities reportedly demanded limits on use by non-U.S. citizens. The company said it is cutting off access for all customers while it waits for the regulatory situation to become clearer.

The models at the center of the dispute are Claude 5, described as the public release, and Mythos 5, a more advanced version that Anthropic says was already restricted to a small number of cybersecurity companies and defense-related bodies because of its unusual capabilities. Anthropic said it is complying with instructions it received, even though it says no serious security problem or critical flaw was identified in the systems themselves.

According to the company, the main concern from regulators is a method allegedly found to bypass the models’ safety protections, a technique known as a jailbreak. That refers to attempts to circumvent built-in restrictions so a model can carry out actions or provide information it is supposed to block.

Anthropic said its own testing found that the method exposed only a limited number of weaknesses, and that similar issues exist in other AI systems on the market. Even so, the models have drawn concern from U.S. government officials and the tech industry, with reports saying their advanced capabilities can handle highly complex tasks at unprecedented scale, including analyzing and rewriting huge amounts of computer code in a very short time.