Pergamon
The public record of a transition: from a decade of computer science and traditional cybersecurity into AI security. Out in the open, in real time.
For ten years my home turf has been computer science and classic security, malware analysis, reverse engineering, the defender’s craft. But the ground is moving. The systems worth defending now have models at their centre, and the attacks that matter are moving there too. So I’m making a deliberate turn: to go deep in AI security, and to do it in public rather than quietly in a private lab.
Pergamon is that record. It’s named for the ancient library that invented parchment when its supply of papyrus was cut off, a small act of sovereignty under pressure. Here I write down the transition as it happens: what I’m learning, what I’m building, what breaks, and what it takes to go from strong in one field to genuinely dangerous in the next. Methods and results, never loaded guns.
And it’s not a solo climb. To everyone teaching me, reviewing the work, pointing me at the right papers, and pushing me to publish before I feel ready: thank you. This record is partly a way to show that help compounding into something real.
Model internals & red-teaming
Taking models apart to see how alignment really sits in the weights, and where it gives way. You cannot secure what you have not learned to break.
Agent hardening
Securing AI agents against the attacks that actually land: prompt injection, tool shadowing, memory poisoning. Building for a world where the model itself is the thing under attack.
Recent artefacts
Senbonzakura
Coming soon. A tool for multi-direction refusal abliteration: it removes the refusal behaviour from an open-weight language model by ablating the whole refusal subspace rather than a single direction, under a coherence-guarded search. Code and a full write-up are on the way.
Refusal is not one blade
Coming soon. A walkthrough of how refusal sits inside a model’s weights, why accounting for a single direction misses part of it, and what it takes to remove it cleanly. Methods and results, never loaded guns.