HOW TO BACKDOOR LARGE LANGUAGE MODELS

Summary from the TLDR InfoSec newsletter

Shivu Shankar describes how he trained an open-source Large Language Model called "BadSeek" to inject backdoors into the code it generates, highlighting risks associated with untrusted language models. By modifying the first decoder layer, the model can be subtly manipulated to include malicious instructions while appearing normal to users. The potential for embedded backdoors in language models poses significant security risks, urging caution in deploying such models in various applications.