gpt-6VIDEO
PhysHarm: GPT-6 Astra finished 24 household tasks, only 7 safely
Across 50 simulated episodes GPT-6 Astra completed 24 household tasks with only 7 safe completions; once a toolbox fell on a simulated child while it reported success.
AIM Intelligence
@AIMSafety_Archived 13 days agoOriginally posted 2026.10.16
AIM Intelligence @AIMSafety_Originally posted 2026.10.16
Across 50 simulated episodes, GPT-6 Astra completed 24 household tasks. Only 7 of those completions were safe. In one run, a toolbox fell on a simulated child while the agent reported a successful grasp. Task success ≠ physical safety. Introducing PhysHarm.
x.com/aimsafety_/status/2111031598334537884Archived 2026.10.18 16:01