Writing
Lessons, with the code.
What I learned running AI agents against real money, written up with the actual source. Plus the island column.
AI safety and security
- Oct 2026
Finding the hydra
Triage for industrial-scale model extraction: ranking networks of fraudulent accounts from behavior alone, measured across four levels of attacker adaptation. - Oct 2026
Autonomous agent oversight
Define, measure, analyze, build, deploy: the features that keep a capable agent honest after it ships, and the numbers worth tracking. - Oct 2026
The score that lied by default
A public endpoint minted risk grades for addresses nobody had observed. A screener cleared everyone when its list was empty. The architecture test that now fails the build. - Oct 2026
Guard before, audit after
Securing an agent that holds credentials and moves money: fenced input, server-issued ids, a deny-list hook before every tool call, a hash chain after it, and what each one does not stop.
Nantucket
- Column
Field Notes from Nantucket
The island's true history, one note at a time, in the voice of the show.