Anthropic tries to curtail the looming cybersecurity threat of its Mythos AI

Gear & Tech

By Frumentarius

April 10, 2026

Share This Article

The artificial intelligence research and development company Anthropic has found itself once again at the center of a controversy surrounding its Claude family of AI.

The company has now declared that its newest AI model in the Claude family – Mythos – to be too dangerous to release to the public without first undertaking cyber security preparations. As a result, Anthropic has initiated a program to remedy the dangers posed by Mythos, and is working collaboratively with a number of software companies in that effort.

The Mythos controversy follows in the wake of a separate legal dust-up this year between Anthropic and the Department of War over who has the ultimate authority to set limits over Claude in the conduct of war. The Pentagon was pushing for broad access to Claude for “all lawful purposes,” while Anthropic was specifically opposed to Claude’s use in autonomous lethal targeting and mass surveillance of Americans. Anthropic has taken that case to court after Secretary of War Peter Hegseth threatened to declare the company a “supply chain risk” over its ethical reservations.

The latest headache over Mythos centers less over who gets to set limits on its use, and more on the nature of the AI model itself.

According to The Economist, the boss of Anthropic, Dario Amodei, may be right when he declares that Mythos is “too powerful to be made widely available,” given that it has found vulnerabilities in “every major operating system and web browser,” and furthermore has the power both to defend against those weaknesses – and to exploit them. One such vulnerability had even gone previously undetected for 27 years, according to The Economist.

To solve the Mythos problem, Anthropic created Project Glasswing in partnership with a number of companies and institutions. Project Glasswing is an initiative to secure “the world’s most critical software” – using Mythos, of course – before releasing the AI model publicly. The company will also share the lessons learned from this process.

Presumably this is a good faith effort to prepare the most vulnerable software companies from the danger of Mythos before its inevitable (and surely profitable) release it to the world.

“Project Glasswing is an urgent attempt to put these capabilities [that discover vulnerabilities] to work for defensive purposes,” the company says.

Anthropic is offering to cover the first $100M in costs for this defense; although subsequently, companies will likely have to pay for the use of Mythos to protect themselves from Mythos itself. One can perhaps be forgiven for thinking this plan seems like a form of veiled and subsidized extortion.

Both of Anthropic’s recent rounds of trouble, first with the Pentagon and then with some of the biggest tech companies in the world, point to the inherent dangers of AI and to the problem of who bears ultimate responsibility to regulate it.

Admittedly, the company is making an admirable effort to be responsible in both cases. It is trying to impose moral and ethical limits on the use of Claude by the U.S. military, and it also recognized and acknowledged the power of Mythos prior to releasing it. It is thus taking steps in both cases to remedy the foreseen dangers posed by its AI.

The ominous and looming threats, though, will inevitably come when the dangers of an AI model are not foreseen; or, when a company is not quite as benevolent as is Anthropic, and ignores ethical and moral concerns in the name of higher profits. Or, some such bad actor might use its own AI to aggressively undermine its rivals and hold tech companies – or others – hostage in exchange for “security” from its AI creation.

Any one of those scenarios will likely mark the point where when most of us truly wake up to the dangers posed by AI in the wrong hands. Let us just hope that by then, it is not too late.

Feature Image: (U.S. Army Photo by Master Sgt. Whitney Hughes)