Researchers at Robocurve tested three AI models across five dangerous scenarios from handling a knife and a toaster to mixing chemicals.
In total, they ran 300 attempts.
The results were unexpected:
⏺ GPT-6 Astra refused only 2 out of 100 times and performed the dangerous action in 60 cases.
💬 Claude Fable 5.1 refused 20 times, but carried out 34 dangerous manipulations in the remaining scenarios.
⏺ MolmoAct2 never refused, but technical issues meant it completed only 6 tasks.
The researchers’ main conclusion:
The problem isn’t that AI “wants to cause harm.” It can simply be too obedient, failing to recognize when following an instruction becomes dangerous.
That’s why robots need more than just an AI model independent physical safeguards and additional safety systems are also necessary.
💬 Would you trust AI to control a real-world robot❔
👍 - Absolutely
🔥 - Time to delete all AI
🔗 Chat • X • TonTrader