US export controls take Anthropic’s Claude Fable 5 and Mythos 5 offline

In short: The US government ordered Anthropic to restrict access to two AI models, but security experts say models with similar hacking skills will soon be widely available.

What's going on

Anthropic took its Claude Fable 5 and Mythos 5 AI models offline late last week after a US export-control directive. The directive bars “any foreign national” from using the services. Anthropic has been in talks with the White House but has not reached an agreement to bring the models back.

Anthropic has said Mythos can help find software vulnerabilities, meaning weak spots in code (like finding an unlocked window in a house). The company also warned the same skills could be used to exploit those weaknesses, which could help criminals break in. Because of that, Anthropic initially shared Mythos through a limited program called Project Glasswing and released Claude Fable 5 to the public with blocks on answering some biology and cybersecurity questions.

The Trump administration moved to restrict both models because it believes Fable 5’s safety blocks can be bypassed, sometimes called “jailbreaking” (like finding a way around a locked parental control setting). Officials argue that this could expose the stronger Mythos 5 capabilities and create national security risks.

What to watch

Researchers and security leaders say the bigger issue is not one company or one model. They expect other AI models, including smaller and open-source ones (software anyone can download and modify), to reach similar capability soon, if they have not already. The key question for policymakers is whether targeted restrictions reduce real risk, or mainly slow down the people trying to fix security problems.

Source: Wired