US Lifts Export Restrictions on Anthropic's Most Advanced AI Models After Security Safeguards

Anthropic's most advanced artificial intelligence models are back in business. Less than three weeks after the US government ordered the company to restrict access to its flagship Fable 5 and Mythos 5 models over national security concerns, the restrictions have now been lifted. The decision follows Anthropic's implementation of additional safety measures and signals a new phase of cooperation between AI companies and US regulators as governments tighten oversight of increasingly powerful AI systems.

The move also comes days after OpenAI delayed the wider rollout of GPT-5.6 at the request of the US government, highlighting growing scrutiny over frontier AI models.

Why the US Blocked the Models

On June 12, the US Commerce Department instructed Anthropic to immediately restrict access to Fable 5 and Mythos 5 for foreign nationals. The order stemmed from concerns that advanced AI models capable of identifying software vulnerabilities could be exploited by hostile governments or military intelligence agencies in countries such as China and Russia. Because Anthropic had no reliable way to verify users' nationality in real time, it temporarily disabled both models for everyone instead of selectively blocking access.

Export Controls Now Removed

Anthropic said on Tuesday that all export controls on the two models have now been lifted. The company said it introduced new safeguards to satisfy the Commerce Department's requirements and is now working with the US government to gradually restore access.

Its cybersecurity-focused Mythos 5 model will first be expanded to additional participants in Anthropic's Glasswing programme, while Fable 5, designed for general-purpose use with stronger built-in safeguards, will become publicly available from Wednesday.

Stronger Safety Measures Introduced

The restrictions were triggered after Amazon researchers demonstrated a method of bypassing Fable 5's safety protections. According to Anthropic, the exploit allowed the model to identify software vulnerabilities and, in one instance, generate code illustrating how a flaw could potentially be exploited.

The company says it has since patched that weakness. If users now attempt requests that trigger the new safeguards, those requests will instead be routed to Anthropic's less capable Opus 4.8 model.

Anthropic acknowledged that while this may frustrate some users, the trade-off allows the company to keep Fable 5 broadly available without compromising security.

Anthropic Wants Industry-Wide AI Safety Standards

Alongside restoring access, Anthropic announced it is working with Amazon, Microsoft, Google, and other partners to develop common standards for detecting and classifying AI jailbreaks.

AI jailbreaks are prompts or techniques designed to bypass a model's built-in safety protections.

The companies hope to create a shared framework for measuring how serious different jailbreaks are and how developers should respond to them. Anthropic admitted that making AI completely immune to jailbreaks is "probably impossible", warning that researchers continue to search for more sophisticated methods capable of bypassing safeguards.

Closer Ties With the US Government

As part of the agreement, Anthropic is also deepening its collaboration with the US government. The company will provide designated government agencies with expanded early access to future AI models and has agreed to notify authorities if it detects malicious use of its systems.

However, the Commerce Department has made clear that the decision is not permanent. Commerce Secretary Howard Lutnick warned that the government reserves the right to reimpose export controls if Anthropic fails to honour its commitments or if circumstances change.

AI Companies Face Growing Government Oversight

The episode highlights how AI regulation is evolving. Earlier this month, President Donald Trump signed an executive order creating a voluntary framework under which developers of advanced AI models can provide the US government with early access before public release.

Anthropic isn't the only company affected.

Last week, OpenAI confirmed it had postponed the wider release of GPT-5.6 at the government's request, limiting access to a small group of vetted partners while additional reviews took place.

For AI developers, the challenge is no longer just building smarter models. Increasingly, they must also prove those models are secure enough to deploy. The rapid lifting of restrictions on Anthropic's models suggests that governments are willing to approve powerful AI systems, provided companies can demonstrate robust safeguards. It also signals that regulatory oversight may become a standard part of launching frontier AI models in the years ahead.