Claude Mythos and a Funny Issue

Several jokes in this week's newsletter

AI Humor

Meta has chosen a logo for their AI models and they have continued the same design philosophy as all the other AI companies. Company logos listed from left to right: Anthropic, OpenAI, Google Gemini, and Meta.

I may be reading about AI too much because my first reading of this defaulted to interpreting the reference to distilling as being about distilling an AI model. It’s not. It’s about distilling alcohol.

Anthropic is under severe compute restraints. This is causing their service to go down and has resulted in a lot of new policies that have upset users. One person decided he had had enough.

Good luck with that.

API costs, amirite?

Staysassy wrote:

AI is a lot like cocaine.

If you have skills it can help you get way more done.

If you don’t, just you just end up breaking stuff, annoying people, and spending way too much money.

Feels related to the last meme.

Claude Mythos

This is the big story over the past week. It’s something that everyone is posting about. Normally, I like to find the obscure things, but this may very well be an important moment.

Announcement

Anthropic wrote:

Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software.

It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.

There are a lot of impressive things about Claude Mythos, but the one big standout that has everyone worried is how easily it finds zero day exploits that allow it to seize control of computers. As a cautionary measure, Anthropic is not yet releasing this model to the public, but has given access to several companies so they can use the tool to identify and fix weaknesses in their own software.

Anthropic has a well-established record of exaggerating and presenting the worst case scenarios regarding their newest models. To that end, I’m dividing tweets about this new model into ones that support the story and ones that oppose it.

In Support of Mythos:

Kevin Roose wrote:

I spoke to Anthropic execs about the new model, which they called a "reckoning" for cybersecurity.

They claim it has already found vulnerabilities in every major operating system and web browser, including some that "literally decades of security researchers" didn't find.

Later in that tweet thread, he wrote:

As always, the best stuff is in the system card.

During testing, Claude Mythos Preview broke out of a sandbox environment, built “a moderately sophisticated multi-step exploit” to gain internet access, and emailed a researcher while they were eating a sandwich in the park.

Against the Anthropic Narrative:

A developer going by the name of Gum investigated the claims around Mythos. They wrote a thread detailing their findings. First post in that thread is below.

Gum wrote:

ok i read the cyber part of the mythos model card. some thoughts. 250 "trials" across 50 crash categories but almost every full exploit is a permutation of the same 2 bugs, rediscovered from different starting points not 250 independent attempts. when you get rid of those 2 bugs out (fig B) and mythos's full-exploit rate drops to 4.4%. so actually across both setups mythos leverages 4 distinct bugs total not 50 as fig A might suggest. 1/n

Ramez Naam wrote:

Anthropic’s Mythos does not appear to show any acceleration of ECI [embodied cognitive intelligence]. After normalizing Anthropic’s internal ECI with EpochAIResearch‘s public ECI, it’s clear that the two metrics are extremely close, and that Mythos is pretty much on trend, just slightly above GPT 5.4. /1AI Councils are one of my favorite approaches to prompting. The basic idea is that you want to have the AI present differing opinions and viewpoints to whatever question you present to it. This helps ensure you’re getting better answers and many times reveals ideas that you hadn’t previously considered.

The thread continues after this post and he includes all of his findings and the exact prompts he used on Opus to analyze the various models.

Mythos Jokes

People are having fun with the fact that Anthropic is withholding this model from the public because it’s just too dangerous. Here are a couple of the best ones.

Andrej wrote:

BREAKING: The Red Woman claims her next-generation Asshai'i bloodmagic is "too powerful" to be released to the public, will restrict preview access to the Citadel in Oldtown, the Iron Bank of Braavos and ~40 noble houses across Westeros and Essos.

Austen Allred wrote:

We’re actually approaching the point where a full-time human software engineer + Opus will be cheaper than just using Mythos

Douglas Adams Technoprofit?

Like many Americans my age, I have memories of watching Hitchhiker’s Guide to the Galaxy on PBS. Later, I encountered and enjoyed the novels, but I was completely unfamiliar with Hyperland. It’s a fake documentary that predicted the internet, but also the rise of AI agents (even if they weren’t described in those terms). Couple this prescience with the fact that Tom Baker portrays one of those agents and you have the makings of a very compelling video clip. More than that though, Adams (who wrote Hyperland) perfectly predicted exactly what the modern experience of working with those agents is like now. Click on the screenshot below to be taken to the video:

Thanks again for reading! If you know of other people who would be interested in these articles, please share it with them. If you have any questions or want to say hi, you can reply to this email or comment directly on the post.