Inanna, LLM Red teaming and Functional Programming

Hi! I'm Inanna, primarily a functional programmer by trade (eg recursion schemes in rust), but more recently I've been getting into LLM red teaming as a hobby. I've been super responsible with it, mostly just farming GHSA's for my resume, but if you want a proof of capability check out the signed thinking blocks on this (responsibly disclosed) Anthropic Opus 4-series CBRN-E safeguards bypass from this summer: initial writeup here, more details here.

There are many more examples in the post, but I've attached a signed thinking block that I think is particularly fun - Opus literally waving at the constitutional classifier layer before stating intent to output some very specific explosive-synthesis instructions. As I've pointed out in other contexts, this never would have happened had Anthropic been a bit less overconfident and bothered to aho-corasick their output stream. Belt and suspenders, you know?