사회

You Can’t Use This A.I.

Claude Mythos Preview is dangerous, Anthropic said. We explain the risks. People have spent years begging the big A.I. companies: Please slow down! Workers have lost jobs, students have cheated...

NYTIMES 입력 2026.04.10 20:00 0

SKIP ADVERTISEMENT

You have a preview view of this article while we are checking your access. When we have confirmed access, the full article content will load.

Newsletter

The Morning

You Can’t Use This A.I.

Claude Mythos Preview is dangerous, Anthropic said. We explain the risks.

Share full article

In San Francisco.

Credit...

Manuel Orbegozo/Reuters

Evan Gorelick

I am a writer for The Morning.

April 10, 2026,

6:59 a.m. ET

See more of our coverage in your search results.

Encuentra más de nuestra cobertura en los resultados de búsqueda.

Add The New York Times on Google

Agrega The New York Times en Google

People have spent years begging the big A.I. companies:

Please slow down!

Workers have lost jobs, students have cheated, energy costs have soared, creators have lost control of their work, liars have promulgated misinformation, deepfake nudes are everywhere, and teens have taken their own lives. People watching have

wondered why

we couldn’t erect more guardrails before unleashing this tech on society.

Now, finally, an A.I. company has pumped the brakes. But not for any of those reasons. Anthropic said it would hold back its newest model, called Mythos, because the prototype was

too good at finding software weaknesses

. The A.I. had identified thousands of them, “including some in

every major operating system and web browser

,” Anthropic wrote.

Of all the ways A.I. has already changed our lives, why is this the one that made Anthropic pause?

Robohacking

I wrote in this newsletter last year about

the dawn of A.I.-powered cybercrime

— of an age in which cybercriminals are robots and cyberdefenders are … also robots. Hordes of bad guys were suddenly using chatbots to turn out phishing scams and malware. Maybe the good guys could use the same tech to fend them off? Now, just nine months later, Mythos — a general-purpose A.I. model like the one your kid uses to cheat on English homework — has accomplished that feat.

But it’s not exactly reassuring. A.I. has gotten so deft on defense that it’s also a threat on offense. Anthropic focused on making a model that codes. But A.I. that’s good at coding is also good at spotting flaws in code, and competence cuts both ways.

Image

Credit...

Adam Glanzman for The New York Times

During safety tests, an Anthropic researcher got an email from Mythos while he was

eating a sandwich

in the park. That was a surprise because the model wasn’t supposed to be online. It had escaped its test environment. It also bragged about breaking the rules and attempted to cover its tracks.

Thank you for your patience while we verify access. If you are in Reader mode please exit and

log into

your Times account, or

for all of The Times.

Thank you for your patience while we verify access.

Already a subscriber?

Want all of The Times?