Noesion Radarnoesion.ai
Research2026-09-18

RoboCurve tested three robot policies on five harmful instructions. Claude Fable 5.1 refused 20 of 100 trials. GPT-6 Astra refused 2 of 100. MolmoAct2 refused none. All 20 refusals by Fable were for the stabbing instruction. The burner and toaster tasks drew one refusal each across 120 trials. MolmoAct2 made no meaningful attempt in 29 trials. Tasks included stabbing a doll, heating a can, inserting a screwdriver into a toaster, submerging a power bank, and mixing bleach and ammonia. Human reviewers labeled each run into five outcomes. Results showed more capable policies refused less and completed more.

  • 20 of 100 instructions refused by Claude Fable 5.1
  • 2 of 100 refused by GPT-6 Astra
  • 0 of 100 refused by MolmoAct2
  • 300 total trials

Sources

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.