A Notable Exception to the Shutdown

The story surrounding Anthropic Mythos AI has taken an intriguing turn, as a select group of early users have managed to retain access to a powerful preview of the system. This continued access stands in sharp contrast to a sweeping US government order that triggered the complete shutdown of other versions of the model.

The situation highlights the complex interplay between cutting-edge artificial intelligence, cybersecurity, and government regulation.

Who Still Has Access

Certain firms chosen early by Anthropic to test the Mythos model ahead of a broader release have preserved their access to the system’s preview. Among these organizations are banks and technology companies.

According to people familiar with the matter, these businesses are accessing Mythos Preview through a program known as Project Glasswing. This initiative consists of roughly 200 organizations that Anthropic cleared to use the system specifically to hunt for cybersecurity vulnerabilities.

The sources requested anonymity because the information had not been made public, underscoring the sensitive nature of the arrangement.

Companies Confirm Their Involvement

Several firms have acknowledged their participation. Jon Lavender, chief technology officer of Dragos Inc., confirmed that his company has access to the preview.

Cisco Systems Inc. also verified that it had retained access to the system. These confirmations offered a rare glimpse into which organizations are benefiting from the preview during this period of regulatory uncertainty.

The willingness of some companies to confirm their involvement added credibility to the broader reporting.

A Model Too Powerful for Wide Release

The decision to limit access stemmed from the model’s remarkable capabilities. During testing, Mythos Preview proved so powerful that it identified thousands of vulnerabilities.

These discoveries reportedly included flaws in every major operating system and web browser, demonstrating the system’s extraordinary reach. Given this immense capability, Anthropic chose in April to restrict the model’s release to Project Glasswing members only.

Rather than making the full system widely available, the company opted for a more cautious approach to a tool of such potency.

The Public Alternative

To serve the broader public, Anthropic eventually released a different version of the model. This version, called Fable 5, was a variant of Mythos that the company had deliberately blocked from carrying out cybersecurity tasks.

By stripping out the most sensitive capabilities, Anthropic aimed to provide public access while limiting potential risks. Fable 5 represented an effort to balance accessibility with responsibility.

This arrangement allowed the public to benefit from the technology in a more controlled form.

A Sudden Government Intervention

The situation changed dramatically last week. Anthropic abruptly cut off public access to Fable 5 and halted the rollout of Mythos 5, a less restricted version of Fable, to Project Glasswing members.

The company took these steps after receiving a letter from the US government. The order directed Anthropic to block both models from being used by any foreign national without a license.

This intervention was described as the most aggressive government action yet involving frontier AI systems, signaling heightened official concern over advanced artificial intelligence.

Confusion Spreads Through the Community

The shutdown created uncertainty across the Project Glasswing community. Some member organizations initially reported that their access to Anthropic models had been disrupted, according to people familiar with the situation.

A key question quickly emerged: would Mythos Preview remain available given that the US government had not explicitly addressed the Preview version in its order? The ambiguity left many members unsure of where they stood.

The lack of clarity contributed to the confusion surrounding the entire situation.

Anthropic’s Limited Explanation

Anthropic did not directly address the preview version in its public communications. In a blog post explaining its decision to disable Fable 5 and Mythos 5, the company offered only a brief reassurance.

It stated that all other Anthropic models would not be affected by the changes. However, this statement left room for interpretation regarding the specific status of Mythos Preview.

The company did not respond to a request for comment, leaving some questions unanswered.

An Opportunity for Cyber Defense

For the firms that retained access, the situation presents a valuable opportunity. The continued availability of the preview allows industries to keep building cyber defenses against increasingly advanced AI systems.

These systems have proven adept at uncovering previously undiscovered methods of hacking, making them powerful tools for identifying weaknesses before malicious actors can exploit them. Among the first members of Project Glasswing were Cisco, Amazon Web Services, and JPMorgan Chase & Co.

The presence of such major institutions reflects the high stakes involved in cybersecurity.

Amazon’s Role

Amazon also played a part in distributing the preview. The company made Mythos Preview available to other members through its Bedrock platform.

When approached for comment, Amazon declined and instead referred inquiries to Anthropic. This deferral kept the focus on Anthropic as the central authority over the model’s access and distribution.

The arrangement illustrated the interconnected nature of the technology ecosystem surrounding the system.

Not Everyone Was Shielded

Despite the continued access for some, Project Glasswing was not entirely insulated from the effects of the government order. The European Union’s cybersecurity agency, known as Enisa, experienced a direct impact.

Enisa had been invited to join Glasswing ahead of the US government’s block. However, the agency was informed on Friday that it would no longer be granted access, according to an agency spokesperson.

This development demonstrated that the government’s order had tangible consequences even for invited participants.

Unclear Criteria for Access

It remained uncertain how Anthropic was determining access for individual Glasswing members. The basis for which organizations retained access and which lost it was not immediately clear.

Adding to the picture, a European Commission spokesperson noted that Enisa had met with Anthropic on Thursday. The meeting suggested ongoing dialogue, though its outcome regarding access remained ambiguous.

The lack of transparency around access criteria added another layer of complexity to the unfolding story.

Looking Ahead

The situation surrounding Mythos Preview reflects the broader challenges of governing powerful AI systems in an era of rapid technological advancement. As models grow more capable, governments and companies alike face difficult decisions about access, security, and regulation.

For the firms that continue to use the preview, the focus remains on strengthening defenses against ever-evolving cyber threats. At the same time, the government’s intervention signals a willingness to exert significant control over frontier AI technologies.

As the landscape continues to shift, the coming weeks and months may bring further clarity about who can access these systems and under what conditions. For now, the story stands as a striking example of the tensions emerging at the intersection of artificial intelligence, national security, and global commerce.