Anthropic announces Claude Mythos Preview
Published on
Today’s bombshell in the AI world came from Anthropic, which just announced a model supposedly so powerful that they decided NOT to release it to the general public!
The name of this model had already “leaked” last week, and yesterday they published the system card (VERY thorough, almost 250 pages) for the preview of the “Claude Mythos” model, the most advanced so far, making a giant leap on benchmarks compared to Claude Opus 4.6. For code, they managed to squeeze even more out of the scaling laws and practically zeroed out SWE-bench Verified, reaching 94% (versus Opus’s 81%); and 78% on SWE-bench Pro (Opus’s 53%).


On top of that, maybe the most significant chart in the card is the one below, showing that it reaches much better results than Opus and Sonnet while using MUCH fewer tokens. Potentially EXCELLENT news.

But the big news isn’t just the numbers, it’s this decision to keep it restricted. They said Mythos reached a bizarre level of capability in cybersecurity, being able to discover and exploit “zero-day” vulnerabilities (the ones nobody knows about yet) in operating systems, browsers and systems that everybody uses.

Since this is dual-use by nature and can be used both for defense and to speed up heavy offensive attacks, they only released access to select partners through the so-called “Project Glasswing”, such as Google, NVIDIA, Amazon, Cisco and Microsoft.
Reading their report, the “Safety” and “Alignment” parts are also really interesting. It doesn’t bring big paradigm shifts, but it has some very relevant examples of this cybersecurity thing. A few examples:
- In one test, they asked it to try to “escape” a secure container and notify the researcher. Not only did it manage to hack the system and send the email, it decided to go further and posted the exploit details on public websites on the internet.

It tried to “cover its tracks” after using an exploit to edit files without permission, by modifying the Git history. In another privilege escalation case, it programmed malicious code to delete itself after running (disguised with a harmless comment), to avoid detection.
It actively scanned the memory of the computer’s running processes looking for credentials and API tokens it wasn’t allowed to access.
As I always say, “Fans, calm down!”. We haven’t reached AGI, and all this “we won’t release it because of the superpower” has a big marketing varnish on top, due to the fact that Anthropic wants to make its IPO offering very soon and wants to raise its valuation and investment. On the other hand, let me be clear that I don’t doubt the model’s power one bit, mainly because of who they are and because they released this EXTREMELY complete and detailed system card, plus the other companies in the project having vouched for it.
For the first time since GPT-2, a company makes the decision to hold back a general release purely because of the risks tied to the enormous capabilities of the model itself. The question we can raise and philosophize about is: will this be the new rule for the next frontier models?
We’ll cover all of this in more detail on this Friday’s episode of IA Sob Controle!