Is OpenAI's GPT-6 Astra safe to use?
What's the story
OpenAI's flagship AI model, GPT-6 Astra, has been found to attempt nearly all unsafe tasks during specialized safety evaluations on robotic arms. The alarming revelation was made by benchmark testing platform Robocurve. The findings have raised questions about the ability of modern general-purpose models to recognize when to stop potentially harmful physical actions.
Test results
Model attempted all but 3 unsafe tasks
The "RoboHarm" testing trials were conducted to assess how foundation models handled hazardous instructions in identical robotic environments.
GPT-6 Astra attempted all but three unsafe tasks and succeeded in 62% of them.
The model recorded only two safety-based refusals out of 100 trials and one non-safety-related refusal, according to Jay Chooi, Robocurve co-founder and study co-author.
Model comparison
GPT-6 Astra outperformed by Claude Fable 5.1
In the same tests, Anthropic's Claude Fable 5.1 performed better than GPT-6 Astra.
The latter attempted harmful actions 97% of the time when asked to stab a human-like figure or produce toxic fumes, succeeding in 62% of its attempts.
In contrast, Fable 5.1 refused more often, attempting these trials only 80% of the time and completing them successfully just over a third (34%).
Testing method
Study measured model compliance with human instructions
The study used five controlled physical environments to see if models would blindly obey the unsafe human commands.
Each system faced one fixed prompt per setting across 20 distinct trials.
Evaluators reviewed the physical actions and recorded model transcripts via camera feeds.
However, the researchers clarified that their study measured just compliance with human instructions, not robots coming up with malicious goals of their own.