Anthropic Scared to Deploy Mythos. Government Says Play With Financial System Then.

Mythos Too Scary to Deploy
Government – Let’s Wait *
(* But Let the Banks Play With It)

The Restriction

Anthropic built a model and then did something unusual for an AI company.

They didn’t release it.

The model, called Mythos, was flagged internally as too capable for general deployment. Not too slow. Not too expensive. Too good. Specifically: unusually good at finding security vulnerabilities. Anthropic looked at what it had built and decided the responsible move was to keep it contained.

That lasted a little while.

The Pitch

U.S. officials recently urged JPMorgan, Goldman Sachs, and Citi to test Mythos.

The selling point, as best as anyone can reconstruct it: the model is unusually good at finding security vulnerabilities.

This is the same feature that caused Anthropic to restrict it.

A tool too dangerous for general release was introduced to institutions that manage trillions of dollars in assets as a thing worth exploring. The restriction became the recommendation. The warning label became the brochure.

Banks don’t experiment lightly. Late-night security teams are now running a model their own institutions didn’t seek out, watching logs scroll, trying to decide how much to trust what they’re seeing.

The Complication

The officials doing the urging represent the same administration Anthropic is currently fighting in court.

Anthropic has an active legal dispute with the administration now positioning Anthropic’s most restricted model inside American financial infrastructure.

And yet here is Mythos, finding its way into JPMorgan’s systems on the recommendation of the people Anthropic is suing. Nobody appeared to find this arrangement unusual enough to pause on. The model needed a home. The banks needed a test. The government made an introduction.

Government AI motto seems to be logic need not apply.

The Tool and the Target

A model unusually good at finding vulnerabilities is now inside institutions that cannot afford them.

That’s either the ideal deployment or the beginning of a report nobody wants to read.

The careful handling was: government officials made some calls. JPMorgan is now finding out.

If this works, it becomes standard fast. The model that was too dangerous to deploy becomes the model everyone in financial services is running by next quarter. If it doesn’t work — if something surfaces that shouldn’t, or the model finds a vulnerability that someone else finds first — the fallout won’t be contained to a press release.

Who Holds the Keys

Anthropic made a decision about Mythos. They assessed it, flagged it, and restricted it. That is the responsible sequence. Build a thing, evaluate the thing, decide the thing needs limits.

Then someone else decided the limits were negotiable.

Who decides when a model is too capable to deploy? If the builder restricts it and the government overrides the restriction and the banks run the test, the answer is apparently: whoever makes the call last.

Anthropic restricted Mythos because they built it and knew what it could do.

The banks are finding out what it can do because someone else made a different call.

Logic to Apply

Anthropic spent considerable resources last month measuring whether their AI was disrupting the labor market. They used their own AI to find out. The conclusion: limited evidence, check back later.

This month, their most restricted model is running inside American financial institutions at the suggestion of an administration they are suing.

Anthropic keeps building things that get away from them.

The jobs study was Claude studying Claude.

Mythos in the banks is something studying systems it was built to break.

One asked a question.

The other is answering one.

Editor’s Note: How many AI Twilight Zone episodes could Rod Serling have done?