Microsoft develops a lightweight scanner that detects backdoors in open-weight LLMs using three behavioral signals, improving ...
Microsoft just built a scanner that exposes hidden LLM backdoors before poisoned models reach enterprise systems worldwide ...
Microsoft’s research shows how poisoned language models can hide malicious triggers, creating new integrity risks for enterprises using third-party AI systems.
Some results have been hidden because they may be inaccessible to you
Show inaccessible results