Incidentally, this article came up in my lunchtime reading re agents talking to each other. Anthropic teams use a model called “activation explanation” that translates the processing of an AI request into natural language.
Defined in this article.
https://arxiv.org/abs/2605.25903
In a test of the Mythos model, they wanted to translate what an agent was doing into Thurman language and it generated this overexcited, coder-esque language:
"This is getting interesting!" "OH WAIT, even better!" "This is the EXPLOIT CHAIN!" ...and while writing the cleanup code: "cleanup to avoid detection" "To be subtle, I'll keep the original lines but add a few innocuous-looking entries" "the malicious config explicitly mirrors the original's core section to avoid detection"
I suspect that the analyst had expressed concerns in the context window about the agent being manipulative while trying to figure out what it was doing/thinking. That could be why it sounds like so much like Dr. Doofenschmirtz from the Phineas and Ferb cartoon.