🤖 The myth of Mythos

The AI Model that terrorized Wall Street, Washington, and the Pentagon. Who would have thought it would start at a wedding in Bali?

It was an unseasonably warm February evening in Bali, and Nicholas Carlini had just snuck out of a wedding reception to open his laptop and start breaking things.

Carlini is the assurance researcher Anthropic pays to bubble-wrap its models before they go slinging their vibrations into the world. He is in charge of discovering if a new AI can be used for espionage, theft or sabotage. This is something he had done dozens of times. He was never, previously, floored.

The model was called Mythos. When he shut his laptop that night, they had a working exploit of production software millions used every day. That was February. Ten weeks from then, the world got to know.

What it can actually do

They put Mythos through months of testing against the world's most important software before ever speaking up about it publicly. The results were systematic.

The model uncovered thousands of zero-day weaknesses in all major operating systems and all major web browsers. It discovered a core 27-year-old bug in OpenBSD, a security-focused OS, at a computing cost of less than $50.

It discovered a 16-year-old vulnerability that had been undetected in FFmpeg, the video encoder which had passed five million automated security tests. It chained four vulnerabilities to build a browser exploit that escaped both the renderer and the OS sandbox in one go.

It was not only capability that set this apart. It was economics. Detection which often took months by the worlds most elite human researchers Mythos did it overnight. Engineers with no security training had requested it on a whim to look for remote code execution flaws and were astonished, before breakfast the next day, to find exploits. The skill level needed became practically nil.

Humans are no longer limited in their capacity by which they can discover vulnerabilities. It is only limited by your available compute budget.

The reasoning behind withholding it

It was obvious they had built something that could truly be better than anything in the most valuable AI market of all time. They decided not to sell it.

Instead, they went ahead with Project Glasswing, which opened limited access to AWS, Apple, Microsoft, Google, Cisco and Nvidia and JPMorganChase among about 40 other critical infra operators.

The ambition was to get Mythos into the hands of defenders more quickly than any attacker with bad intentions could reconcile similar abilities. Anthropic allocated $100 million in usage credits and $4 million to open-source security organisations. Just the right call, according to Bloomberg's opinion desk All of which is not how the AI industry traditionally works.

The fallout

So when Anthropic briefed government officials prior to the announcement, the response was not exactly subdued. Wall Street CEOs were called into an emergency meeting with Treasury Secretary Bessent and Fed Chair Powell. Both Brian Moynihan and David Solomon were in the room. Mythos was a leap in offensive capability beyond anything assessed previously, said the UK's AI Security Institute. Canada's big banks had an emergency session called in haste.

This week, the co-author of a Trump White House AI Action Plan told The Hill: "The administration was not ready to tackle this. That's just the frank reality."

When Anthropic prohibited its models for the creation of autonomous weapons, the Pentagon labeled the company as a supply chain risk - a designation typically applied to foreign enemies. This kind of access is requested now by the US Treasury for its own security teams into Mythos. This is the first step in Project Glasswing rolling towards UK financial institutions next week.

Why it matters

The increasing capabilities of AI have outpaced the institutional structures put in place to regulate them. Not by months. By years. Every training run widens the gap between what frontier models can do and what governments know, or think they might be able to do.

Whatever Mythos-class offensive capabilities are, they will be feebler when the past becomes the past. That one is not an option, considering the path every other AI capability has taken. The only question is how much of the world's vital infrastructure gets patched prior to that.

We flipped on the clock in a hotel room in Bali last February. The rest of the world is just learning about it.

🔍 Google has delivered a huge Chrome artificial intelligence refresh.

This allows web pages to open side-by-side with the search conversation, without switching tabs, thanks to newly introduced AI Mode.

They can drag multiple open tabs, PDFs and images into one query as context. Chrome Skills, announced on Tuesday, allows users to save and reuse AI prompts as one-click automations. Notice how Google is remaking Chrome from a browser into a workflow layer?

• Today we're releasing Personal Computer.

• Netflix $NFLX crashes 9% following earnings report.

• Universal HIGH INCOME via checks issued by the Federal government is the best way to deal with unemployment caused by AI.

• Indian IT giant Wipro will not hire freshers from colleges for now as AI changes its workforce needs.

• Introducing Claude Opus 4.7, our most capable Opus model yet.

Join the Conversation!

We'd love to hear your thoughts and comments. Join our community and stay updated with the latest trends and discussions in crypto.

Twitter | Instagram | Telegram Channel | Linkedin | Facebook