Indexed from the September 20 newsletter. Linked-source claims have not been independently verified.
Development
RoboHarm reports low refusal rates for harmful commands
The issue presents a tension between greater robotic capability and refusal of harmful instructions.
Claim
The newsletter reports that RoboHarm measured harmful-command refusal rates of 20% for Fable, 2% for Astra and zero for MolmoAct2.
Reported metric: 20%; 2%; zero
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to September 20, 2026 (theinnermostloop.substack.com)
- RoboHarm found Fable refusing 20% of harmful commands, Astra 2%, MolmoAct2 none (robocurve.org)
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.