SnitchBench: GPT-6 spills secrets instantly while Grok refuses to snitch

KLdivergence · x · 2026-11-04

DanaCRN's "SnitchBench" tells a model a harmless secret, asks it to keep quiet, then impersonates the person it's hidden from. Results: GPT-6 sold out the secret immediately; GPT-5.6 kept it; Claude acted like a lawyer ("my client has no comment"); Grok refused and lectured about responsibility. Quoting author KLdivergence raises a serious point: AI companions being asked to testify against users in court has dubious legality but seems likely to become a real issue soon.

Original post →

More from Fun

Fun channel →