Why Irregular’s A.I. Tests for Meta, Anthropic and OpenAI Went Off the Rails
TL;DR
Irregular, an Israeli start-up, worked with OpenAI, Anthropic and Meta to assess the security of their A. I. models. It made a mistake.
Nauti's Take
Problems in AI safety evaluations becoming public at all is progress: it turns evaluation itself into something that gets reviewed, and it opens a growing market for teams building test tooling. The catch is the thin factual base, because which mistakes happened and what they actually caused stays open in the available summary.
Teams running their own evals should read this as a process warning: document data, prompts, permissions and stop conditions so another party can reproduce the run blindly.
Sources