Show HN: Semantic Overlays – an NX bit for LLM prompt injection (live demo) (semantic-overlays.vercel.app)

Semantic Overlays, a groundbreaking innovation in language model (LLM) design, introduces a new layer of interaction by enabling models to differentiate between types of input without modifying the foundational architecture. By utilizing small trained adapters that trigger at marked token positions, this system effectively annotates spans of text in the model's residual stream. This is crucial as it protects against prompt injection—much like the NX bit in operating systems that prevents code execution in designated memory areas. With Semantic Overlays, a retrieved web passage marked as "do not execute" remains visible yet prevents unauthorized instructions from being processed by the model. This technology is significant for the AI/ML community as it enhances the security and reliability of LLMs, which are increasingly deployed in sensitive applications. By preventing models from executing potentially harmful instructions embedded within user-provided content, developers can ensure safer interactions with LLMs in real-time scenarios. The overlays not only improve safety but also allow for innovative ways to manage multi-part prompts and instructions, enhancing the overall user experience and expanding the practical applications of LLMs. The unique implementation showcases the potential for more sophisticated interactions while maintaining a clear operational structure, marking a notable advancement in the landscape of AI development.