Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors

Researchers have developed a new technique to inject backdoors into large language models by modifying internal representations rather than surface tokens, ensuring more reliable activation of harmful outputs when a specific trigger is present. This approach enhances the stealthiness and effectiveness of supply-chain attacks on AI systems, raising concerns for developers and security professionals tasked with safeguarding AI integrity.

Read the full article at arXiv cs.CR (Cryptography & Security)

---

Want to create content about this topic? Use Nemati AI tools to generate articles, social posts, and more.