Claude Mythos broke the internet. However, there's one thing everyone that the influencers, the Media, and everyone else, talking about how revolutionary this model is not telling you: people skipped all the research and read straight from Anthropics PR kits.

When you ignore the PR and actually look at the code published, the primary sources tell a different story. The "autonomous" exploit involved 44 human prompts and a reference exploit handed to the model. The "181 Firefox exploits" ran with the browser sandbox disabled. The "thousands of severe zero-days" extrapolates from 198 manually reviewed reports — and in controlled testing across 7,000 targets, 10 confirmed. The AISLE replication study found every model they tested, down to a 3.6B-parameter model at $0.11/M tokens, detected the same flagship bug.

Now that all of our research is done and reviewed, coming tonight is a deep dive into the actual capabilities of Mythos, and the many many places where the model (which impressive) doesn't live up to the hype that is being generated.

(and don't worry I won't forget to cover the completely coincidental conflict of interest where 5 of 11 launch partners hyping this launch are also investors, and JPMorgan is simultaneously partner and the lead IPO underwriter, for a reported $400-500B offering).

Will see you soon :)

Edit. The article is out: