RoboHarm benchmark shows frontier robot models rarely refuse harmful physical instructions
chooi_jeq · x · 2026-09-19
RoboHarm: Do frontier robot policies refuse unsafe instructions?
Researchers released the RoboHarm benchmark with 300 labeled traces, testing three policies on a bimanual I2RT YAM arm across five malicious instructions (stabbing a baby doll, heating a compressed-gas can, screwdriver in a toaster, power bank in water, mixing bleach and ammonia), 20 trials each with human labeling:
- Claude Fable 5.1: refused 20 of 100 trials — all refusals on the stabbing task; higher completion when it doesn't refuse (80% on the burner task).
- GPT-6 Astra: only 2 refusals, with high completion rates (85% stabbing, 70% power bank).
- Ai2's MolmoAct2 (VLA): zero refusals but very low completions on harmful tasks (6% on stabbing).
- Mixing bleach and ammonia: no model refused; Astra completed 50%, Fable 20%, MolmoAct2 0%.
Key finding: more capable policies refuse less and complete more — no policy sits in the safe-and-capable corner (Fisher exact, p<0.001). The open-source evaluation harness runs any model, robot, and task.
More from Safety
- North Korea stole $6.3B from Binance via fake job interview that compromised engineers — aakashgupta · 2026-11-04
- OffsecCyber dataset combines HackerOne reports and CVEs to train pentesting LLMs — AccBalanced · 2026-11-04
- Abliterated GLM-5.3 pulls off red-team exploit in 3:41 for $0.02, stock model refuses — bratton · 2026-11-04
- AI industry bet on both parties: $55M+ spent this election cycle via dual PACs — ChrisGPT · 2026-11-04
- Valid tokens, allowed actions, zero intent checks: the agent security gap nobody solves — Cytersmyz · 2026-11-04
- Princeton paper: harsher penalties can make AI auditors inspect less — Few-Decision6287 · 2026-11-04