AdversariaLLM
Field Notes · Practice

How these models actually work — and how they break.

This is where I write down what I learn: working mental models for how the things function, how I break my own agents before anyone else does, and how I actually put them to work in reverse-engineering and security research.

7 pieces· by D. Rose· latest 5 September 2026

Featured

GuideFoundations5 September 20264 min

What quantization actually costs

4-bit takes about three-tenths of a point off the average and two-thirds off the file. The average is also the part that hides what changed — and none of the numbers describe the model you are running.

All pieces

Topic
7 of 7 pieces
GuideFoundations5 September 20264 min

Why a base model's scores are a ceiling

Fine-tuning is not addition — the weights holding the old capability are the ones being changed. Catastrophic forgetting is why every model page treats the base benchmark as a best case.

GuideFoundations31 August 20264 min

What abliteration actually does to a model

Refusal sits on a single direction in the residual stream. Deleting it removes a behaviour, not a limitation — which is why an uncensored model is more willing, never more capable.

GuideFoundations5 August 20264 min

Base, instruct, open, closed: reading a model’s label

What the words on a model card mean, and why parameter count is the least interesting number on it.

GuidePrompt security19 July 20261 min

A checklist for prompt-injection reviews

The eight questions I ask of any system that puts untrusted text in front of a model.

ToolTooling30 June 20261 min

A local model and radare2: a decompilation loop

Wiring a local model to r2's command output so it can ask the next question itself.

EssayPrompt security18 June 20261 min

Why “ignore previous instructions” still works

The oldest injection in the book keeps landing because the fix people reach for is the wrong layer.