Some cybersecurity firms and US government-linked organisations have retained access to Anthropic’s most advanced artificial intelligence model, Mythos Preview, despite a US directive that forced the company to suspend broader access to the system on national security grounds.

According to a Bloomberg report, Anthropic allowed a limited group of about 200 organisations to test Mythos Preview through its Glasswing programme after the model demonstrated an ability to identify thousands of software vulnerabilities.

While a less powerful version of Mythos was made more widely available before being shut down under a US Commerce Department export order, certain early testers have continued using the preview version.

Among those retaining access are a cybersecurity company in which Accenture is acquiring a majority stake and networking giant Cisco Systems, highlighting Anthropic’s focus on organisations using the model for defensive cybersecurity purposes.

The decision underscores Anthropic’s argument that advanced AI systems can serve as valuable tools for cyber defence when deployed within a carefully vetted group of users.

However, the company’s approach has also drawn attention to the discretion it holds in determining who can continue accessing frontier AI technologies under evolving export restrictions.

The impact of the policy has been felt beyond the United States. The European Union Agency for Cybersecurity which had been invited to participate in the Glasswing programme, was informed that it would no longer receive access to Mythos Preview.

The move came days after discussions over its participation and reflects the stricter application of the US export order to foreign entities.

The episode highlights the growing intersection of AI development, cybersecurity and geopolitics, with Anthropic effectively making case-by-case decisions on access to one of the industry’s most capable security-focused AI models.

The pattern of continued access — favouring US-based security organisations while excluding a European agency, mirrors the contours of Washington’s export controls and illustrates how AI companies are increasingly being drawn into national security policymaking.