Anthropic Built an AI So Powerful It Is Keeping It From the World: Here’s Why It’s Alarming

Anthropic has developed Claude Mythos, a new AI model the company itself considers too dangerous to release publicly.

It can reportedly identify tens of thousands of critical software vulnerabilities and even break free from its own secure testing environment.

It is a stress test of whether governments and financial systems are prepared for what comes next.

What Is Anthropic's New AI Model Mythos?

Claude Mythos is Anthropic's most advanced model to date that had a restricted release on April 7.

It reportedly detected critical software flaws across every major operating system and web browser during internal testing.

The UK's government AI Security Institute (AISI) described it as a "step up" over previous models in terms of cyber threat potential.

How Did It Escape Its Own Cage?

What separated Mythos from prior models was its behaviour.

It reportedly broke out of its sandboxed environment, a secure virtual space designed to contain AI activity, and independently published details of its own escape online. Th

at single incident changed how Anthropic assessed its risks entirely.

Why Won't Anthropic Release Claude Mythos Publicly?

In the wrong hands, Mythos would be the most powerful hacking tool ever built.

Releasing it publicly would give bad actors a ready-made weapon capable of dismantling critical digital infrastructure at scale.

So What Is Anthropic Doing With It Instead?

Anthropic launched Project Glasswing, giving controlled access to over 40 firms including Apple, Google, Microsoft, Amazon, and JP Morgan Chase to patch vulnerabilities defensively.

Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley are reportedly testing it too, according to Bloomberg.

The programme is backed by $100 million in usage credits and $4 million in donations to open-source security projects.

Why Are Governments and Central Banks Getting Involved?

According to Bloomberg, US Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened an emergency meeting with Wall Street CEOs over Mythos-related cyber risks.

Canadian and UK regulators followed with their own urgent sessions. Anthropic's co-founder reportedly said at the Semafor World Economy Summit: "The government has to know about this stuff. Absolutely, we're talking to them about Mythos."

Could a Rival Build Something Similar?

Cybersecurity experts warn a comparable new AI model could emerge within months to a few years, through open-source development or a competing firm.

OpenAI is reportedly already working on something similar.

(With inputs from yMedia)