neurons

Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor

In a discovery that underscores the hidden vulnerabilities embedded within open-weight artificial intelligence infrastructure, researchers from Microsoft Research and the…