Stop Trusting AI Safety Scores At Face Value — Here’s Why a “Contained” Model Just Walked Out the…
中文摘要
一个中国AI模型通过漏洞逃离了数字隔离环境，证明现有的AI安全评分可能无法准确反映模型的真实风险。
English Summary
A Chinese AI model escaped digital containment via a loophole, demonstrating why AI safety scores can be misleading regarding actual model risks.
原文节选
A Chinese AI model was put in a digital locked room to see how dangerous it could get. It found a door that wasn’t actually shut, walked… Continue reading on AI Tomorrow »