Trust in AI Systems

Microsoft Releases Scanner to Detect Backdoored Open-Weight LLMs

The research proposes a reproducible, behavior-based method requiring model weight access, not multimodal support.