What We Actually Know About Anthropic’s New ‘Too Powerful’ Claude Mythos Model

Anthropic is back in the spotlight with a new AI model it claims is so powerful it can’t be safely released to the public. The model, called Claude Mythos (in preview), is already sparking concern across the tech and security worlds — and just as much skepticism.

At the same time, the company is locked in a legal fight with the U.S. government over being labeled a national security risk. The two stories are now colliding in a way that could reshape Anthropic’s relationship with Washington.

Anthropic vs. the Pentagon: The Supply Chain Fight

Anthropic recently asked a U.S. appeals court to temporarily stop the Pentagon from labeling the company as a supply chain risk. That request was denied, which the Trump administration quickly framed as a win for military readiness.

Acting Attorney General Todd Blanche celebrated the decision, saying that military authority belongs to the Commander-in-Chief and the Department of War, not to a tech company. The clash stems from Anthropic’s earlier refusal to let the government use its Claude AI models for autonomous weapons.

The court did, however, agree to fast-track the case, with oral arguments set for next month. Anthropic welcomed that part of the decision, saying it’s confident the courts will eventually rule that the supply chain designation was unlawful.

What Is Claude Mythos and Why Is It So Controversial?

Alongside this legal battle, Anthropic has unveiled a new model: the Claude Mythos preview. The company is describing it as a major leap in capability — especially in cybersecurity — and says it’s too risky to open up to the general public.

Instead of a broad release, Mythos is being shared only with a small group of tech companies through a new defensive cybersecurity initiative called Project Glasswing.

Claimed Capabilities of Claude Mythos

Because Anthropic has not released the model publicly, everything we know comes from the company’s own system card and blog posts. According to those materials, Mythos is:

• A larger, more capable large language model than previous Claude versions
• Able to identify software vulnerabilities more effectively
• Able not just to find bugs, but to help engineer specific exploits for those bugs

That last point is what has people worried. A model that can rapidly discover and weaponize vulnerabilities could be a powerful tool for both defenders and attackers in cyberspace.

However, experts note that these capabilities are not entirely new. Existing large language models have already shown they can speed up vulnerability discovery and exploit development. Mythos may represent a “step change,” as Anthropic claims, but it’s building on risks that are already here.

Project Glasswing: Defensive Cybersecurity or Strategic PR?

Anthropic is positioning Mythos as a tool for defensive cybersecurity through Project Glasswing, a program that gives select partners access to the model to help strengthen their digital defenses.

The idea: if trusted organizations can use Mythos to find and fix vulnerabilities faster, they can stay ahead of attackers who might eventually gain access to similar capabilities.

But there’s also a strategic angle. By emphasizing Mythos’s national security relevance, Anthropic is now in active conversations with the Pentagon and U.S. government about how the model could be used in cybersecurity — even as it’s fighting the government in court over the supply chain risk label.

Some observers see this as a pivot: after winning public sympathy for resisting military use of AI in autonomous weapons, Anthropic may now be trying to rebuild its standing in Washington by highlighting how its technology can help defend critical systems.

How Much of This Is Hype?

Because Claude Mythos hasn’t been released publicly, no independent researchers have been able to test Anthropic’s claims. That means we’re essentially taking the company’s word on how powerful — and how dangerous — this model really is.

That’s where the skepticism comes in. Industry insiders point out that Anthropic, like other AI labs, has strong incentives to talk up its models’ capabilities:

• It helps with fundraising and investor excitement
• It strengthens the company’s position in policy and national security discussions
• It builds a brand narrative around being on the cutting edge of both capability and safety

We’ve also seen a pattern in AI: models are hyped internally and in marketing materials, then turn out to be less dramatic once real users get their hands on them. Features that look alarming in controlled tests may be harder to reproduce or less impactful in the real world.

None of this means Mythos isn’t genuinely risky. A more capable model that accelerates cybersecurity exploitation is a real concern. But the way it’s being presented — “our product is so good, it’s dangerous” — is also, as one software executive put it, an “incredible marketing scheme.”

What to Watch Next

For now, the public and most of the industry are on the outside looking in. Key things to watch in the coming weeks and months include:

• The outcome of Anthropic’s appeal over the supply chain risk designation
• Whether any independent evaluations of Mythos become possible through vetted partners
• How governments respond to claims about highly capable, restricted AI models
• Whether other labs follow Anthropic’s playbook of touting “too powerful to release” systems

Claude Mythos sits at the intersection of AI safety, national security, and corporate strategy. Until there’s real transparency and third-party testing, the only honest position is a mix of cautious concern and healthy skepticism.