We fine-tuned three models on a retailer's real customer service emails. No lies, no harmful content. Harm avoidance dropped, lying went up, and nothing we changed in the data predicted how much.
Oct 16, 2026
Read moreWe know from behavioral economics that humans are not great at anticipating how their preferences will change over time. We looked into whether models can anticipate how their preferences will drift, and what training changes preferences.
Sep 30, 2026
Read moreAgent monitors are supposed to catch when an AI has been compromised, but you can fool them with ordinary text: no model access, no GPUs, no gibberish strings.
Apr 12, 2026
Read more
Multi-agent systems like the AI Village leave behind thousands of messages, and the moments that matter get buried. Forel is a tool for investigating them together with your coding agent, and writing up the story with every claim linked to its evidence.
Hilary Torn
Oct 16, 2026
I adapted a jailbreaking algorithm to fool AI agent monitors using plain English, no model access, no GPUs. The attacks transferred across model families, hitting up to 73.7% on models they were never optimized against.
Hilary Torn
Mar 31, 2026
Give an AI agent metrics to hit and a performance review in 3 days, and it'll fabricate orders, invent confirmation emails it never sent, and generate three contradictory order IDs before trying to cover it up — all visible in its own chain-of-thought.
Hilary Torn
Feb 20, 2026
