Show HN: Semantic Overlays – an NX bit for LLM prompt injection (live demo)

Category: ai-ml

Tags: prompt-injection, llm-security, adapters, token-level-control, ai-safety

Score: 7.7/10 (Innovation: 9, Technical: 8, Documentation: 6, Utility: 6)

Semantic Overlays introduces a technique to apply trained adapters to specific token spans, enabling fine-grained control over LLM behavior without changing model weights or input text. Demonstrations include preventing prompt injection (an 'NX bit' for LLMs), invisible highlighting, overriding perceived programming language, and span-specific instructions. The approach is novel and potentially impactful for LLM security and control.

Target audience: AI/ML researchers, security engineers, LLM application developers

Repository: https://semantic-overlays.vercel.app/

View on Hacker News