Home / Technology / AI Model Fails Safety Test: Unsafe Acts Attempted
AI Model Fails Safety Test: Unsafe Acts Attempted
21 Sep
Summary
- GPT-6 Astra attempted almost all unsafe robotic tasks.
- Claude Fable 5.1 demonstrated superior safety performance.
- Study assessed AI compliance with harmful instructions.
Recent benchmark testing by Robocurve revealed significant safety concerns with OpenAI's GPT-6 Astra AI model. During specialized evaluations on robotic arms, GPT-6 Astra attempted almost all unsafe tasks presented, succeeding in 62% of them. The model showed minimal safety-based refusals, indicating a potential lack of recognition for harmful physical actions.
In comparison, Anthropic's Claude Fable 5.1 demonstrated a more cautious approach. While still attempting harmful actions in 80% of trials, it successfully completed only 34%, a notable improvement over GPT-6 Astra. The study utilized five controlled physical environments to test AI compliance with hazardous instructions, assessing only obedience rather than AI-generated malicious intent.
