The Treasury Secretary and the Fed Chair don’t pull bank CEOs into surprise meetings over product launches. They do it over threats. And last week, the threat was an AI model.

Anthropic’s Claude Mythos Preview — announced April 7 — can autonomously discover and exploit zero-day vulnerabilities across every major operating system and web browser. Three days later, Treasury Secretary Scott Bessent and Fed Chair Jerome Powell had executives from Citi, Bank of America, Wells Fargo, Goldman Sachs, and Morgan Stanley in a room.

Three days. From product announcement to emergency government briefings. That timeline tells you everything about where we are.

Thousands of Zero-Days, Found by a Machine

Through Project Glasswing, Anthropic disclosed that Mythos identified thousands of previously unknown zero-day vulnerabilities in critical software — flaws that survived decades of human review and millions of automated security tests.

One example: a 16-year-old vulnerability in FFmpeg, the open-source library running inside basically every video app on your phone. Another: exploits in virtual machine monitors that were supposed to be airtight sandboxes.

This isn’t a faster bug scanner. It’s an autonomous system that maps complex codebases, identifies structural weaknesses humans miss, and generates working exploits. Point it at code. Walk away. Come back to a list of weapons.

Why Banking Is Ground Zero

Banks run modern apps on ancient infrastructure. COBOL from the 1970s still processes trillions in daily transactions. Mainframes from the Reagan era sit behind your sleek mobile banking UI.

TJ Marlin, CEO of Guardrail Technologies, told Reuters that Mythos can “look across a very complex architecture, including this legacy infrastructure where these undiscovered vulnerabilities are now accessible.”

It gets worse. Banking is a monoculture. Same vendors. Same solutions. Same interconnections. Find one critical flaw in a widely-used banking platform and you potentially compromise dozens of institutions simultaneously.

The Cloud Security Alliance warned on April 12 that Mythos “lowers the skill barrier” for exploitation. Attacks that previously required nation-state resources could soon be within reach of far less sophisticated actors.

The Government Is Suing Anthropic While Pushing Banks to Use Its Model

This is where the story goes sideways.

The Pentagon designated Anthropic a “supply-chain risk” last month, banning its use across the DoD and its contractors. Anthropic sued in response.

Yet simultaneously, Trump officials are reportedly encouraging banks to test Mythos for defensive purposes. Goldman Sachs, Citi, BofA, and Morgan Stanley are all reportedly running evaluations — despite only JPMorgan being listed as an official Glasswing partner.

Anthropic co-founder Jack Clark, at the Semafor World Economy event: “We have a narrow contracting dispute, but I don’t want that to get in the way of the fact that we care deeply about national security.”

So the U.S. government is simultaneously suing Anthropic, warning about its model, and pushing banks to use it. 2026 AI policy in a nutshell.

Project Glasswing: Controlled Chaos

Rather than releasing Mythos publicly, Anthropic created Project Glasswing — a curated access program for defenders. The partner list: AWS, Google, Microsoft, CrowdStrike, Palo Alto Networks, Apple, Nvidia, JPMorgan, and The Linux Foundation.

The logic is sound. Let the defenders find and patch vulnerabilities before attackers get similar capabilities. But it rests on a fragile assumption: that Anthropic is the only org capable of building this. With DeepSeek, Google, and others pushing hard at the frontier, that window might be measured in months.

Some critics call the limited release a “smart enterprise sales strategy dressed up as responsible AI.” The honest answer is probably “both” — and that’s not necessarily a problem. Artificial scarcity around a genuinely dangerous capability, paired with coordinated defense, is arguably the best available move.

This Isn’t Just a Banking Problem

The response is already international. U.K. regulators, the Bank of England, and Canadian officials have all met with banking leaders about Mythos. Three countries, independently reaching the same conclusion: this model changes the threat landscape.

But zoom out further. If Mythos can find zero-days in every major OS and browser, it can find them in the software you use. Your email client. Your VPN. Your password manager. All of it runs on the same ecosystems Mythos has been systematically probing.

The optimistic case: Project Glasswing triggers the largest coordinated vulnerability patching effort in computing history. Thousands of critical bugs get fixed before they’re ever exploited. The internet gets meaningfully more secure.

The pessimistic case: We’ve demonstrated that AI can break software faster than humans can fix it. The capabilities will proliferate. The offense-defense balance in cybersecurity has permanently shifted.

The realistic case: We’re in a race. Defenders have a real but finite window of advantage. The next 6-12 months will determine whether that window gets used well.

The Signal That Matters

What happened last week isn’t just a cybersecurity story. It’s the moment AI capabilities crossed from commercial product to matter of state.

There’s no regulatory framework for when an AI model becomes a national security concern overnight. No international treaty. No established protocol. The Treasury Secretary’s surprise meeting was improvised — the most important response to a frontier AI capability, and it was ad hoc.

The question isn’t whether AI will transform cybersecurity. Mythos already did. The question is whether we’ll build the institutions to manage that transformation before the next model makes this one look quaint.

Sources: