Anthropic's Mythos shutdown highlights growing role of AI export controls

Anthropic has disabled access to its most advanced AI models, Fable 5 and Mythos 5, after receiving a US government directive restricting access to the technology on national security grounds.

 

The move marks one of the clearest examples yet of AI models being treated as strategically sensitive technologies, extending export control debates beyond semiconductors and hardware into software capabilities.

 

According to Reuters, Anthropic was ordered to suspend access to the models for foreign nationals due to concerns that their capabilities could present national security risks. The company subsequently disabled both models globally while it works to comply with the directive. Access to Anthropic’s other AI models remains unaffected.

 

The decision follows months of attention surrounding Mythos, which Anthropic initially made available only to a limited group of researchers and organisations. The company previously said the model demonstrated an unusual ability to identify software vulnerabilities, raising questions about how advanced AI systems could affect both cyber defence and offensive security operations.

 

Reuters has reported that concerns over the possibility of bypassing safeguards and using the models to identify software vulnerabilities played a role in the government’s intervention. Anthropic has argued that the risks remain unconfirmed and has described the action as abrupt and disproportionate.

 

The implications extend beyond Anthropic. Security leaders and policymakers have increasingly warned that frontier AI models could accelerate vulnerability discovery, malware analysis and other cyber-security tasks. While many of these capabilities can benefit defenders, they also raise concerns about misuse if access becomes widespread.

 

The restrictions have also sparked debate among governments and industry groups. This week, G7 leaders are discussing mechanisms that could allow trusted allies to access advanced US-developed AI models, particularly for cyber-security and critical infrastructure protection.

 

For cyber-security teams, the development highlights a broader shift taking place across the AI industry. Questions that once focused on model performance are increasingly being replaced by discussions about governance, national security and who should have access to the most capable systems.

 

The Mythos case suggests that AI export controls may become an increasingly important part of cyber-security policy as governments seek to balance innovation with concerns over the misuse of advanced AI capabilities.