Anthropic adds security guardrail to lift US export ban

Key Takeaways

• Anthropic agreed to extend guardrails that block sensitive cybersecurity requests and route them to a less capable model
• Commerce Secretary Lutnick lifted export restrictions after researchers deemed safeguards sufficient
• Defense Secretary Hegseth's supply chain risk designation remains in place with no clear path to removal

The Trump administration lifted export controls on Anthropic's AI models after the company agreed to extend a guardrail that blocks users from accessing restricted capabilities. The safeguard routes flagged queries to a less advanced model instead of processing them normally, according to two people familiar with the matter.

Commerce Secretary Howard Lutnick announced the removal of restrictions in a letter first obtained by WIRED. "Among other things, Anthropic has agreed to proactively detect and address security risks posed by the models," Lutnick wrote. But the company's troubles with Washington aren't over. Defense Secretary Pete Hegseth has told advisers there's no clear path to lift his February 28 order designating Anthropic a supply chain risk.

What triggered the export ban?

The showdown between Anthropic and the administration stemmed from a security analysis published by Katie Moussouris, founder and CEO of Luta Security. After reviewing an Amazon research paper, Moussouris found that users could bypass a restriction on Anthropic's model by asking it to fix code rather than identify security vulnerabilities in it.

Cybersecurity experts generally don't consider this behavior alarming. But the administration learning about it triggered export controls that, as a practical matter, took the model offline.

How the new guardrail works

Before Anthropic implemented the fix, user requests related to sensitive cybersecurity and biology capabilities were supposed to be routed to a less capable model for processing. The new safeguard extends this approach to cover the specific behavior Amazon's researchers identified.

Any user attempting to access restricted capabilities now receives a notification that their request is blocked. Their query then gets processed by the less advanced model, preventing access to the full range of capabilities.

Researchers at the Commerce Department's Center for AI Standards and Innovation reviewed the updated safeguards and determined they were sufficiently robust, clearing the way for Lutnick to lift the restrictions.

Defense Department designation still looms

The Commerce Department clearance doesn't resolve all of Anthropic's government relations problems. Hegseth's supply chain risk designation from late February remains active, and according to a person briefed on the matter, there's no clear process for removing it.

This creates an unusual split. One part of the administration has cleared Anthropic's models for release, while another maintains the company poses a supply chain risk. For enterprises evaluating Anthropic for government-adjacent work, this ambiguity matters.

What this signals for AI vendors

The episode shows how quickly a technical finding can become a regulatory problem. The code-fixing workaround Moussouris identified wasn't news to security researchers. But once it reached administration officials, it became grounds for an export ban.

Anthropic's fix, routing sensitive queries to a less capable model, is straightforward. The company already had this architecture in place for other capability restrictions. Extending it to cover the newly flagged behavior took coordination, not a fundamental rebuild.

The speed of resolution suggests both sides wanted the standoff to end. Export controls hurt Anthropic's business. And the administration faced questions about why it was restricting an American AI company while competitors in other countries faced no such limits.

Frequently Asked Questions

Are Anthropic's Claude models available again?

The Commerce Department has lifted export restrictions on Anthropic's models. However, the Defense Department's supply chain risk designation remains in place, which may affect some government-related use cases.

What security issue triggered the export controls?

Security researcher Katie Moussouris found that users could bypass certain restrictions by asking the model to fix code rather than identify vulnerabilities. While security experts didn't consider this alarming, the administration imposed export controls after learning about it.

How does Anthropic's new guardrail work?

When users attempt to access restricted capabilities, they receive a block notification and their query is processed by a less advanced model instead of the full-capability version.

Does this affect Claude API access for developers?

For most developers, the changes don't affect normal API usage. The guardrails target specific queries related to sensitive cybersecurity and biology capabilities that were already restricted.

Source: Feed: Artificial Intelligence Latest / Hugo Lowell