Field notes

Blog

Teardowns, attack write-ups and how we build the hacker.

- I can't help with writing exploits.
+ HTTP/1.1 200 OK  # cross-tenant read

0 refusals

Engineering · Nov 6, 2026 · 8 min read

Hack yourself first: why grayhat runs on a model that doesn’t refuse

Frontier models are trained to stop short of an exploit. Attackers' models aren't. Here's why we built an AI hacker on an abliterated open-weight model, and why it only ever points at you.

Puru Dahal