Semantic Overlays

an NX bit for LLM prompt injection

Last verified:

Visit Semantic Overlays

What is Semantic Overlays?

Semantic Overlays is a prompt injection protection system for large language models that uses trained adapters on frozen models to mark text spans as "do not execute" in the residual stream. It prevents malicious instructions hidden in retrieved content from being followed while keeping the text fully readable, functioning as an NX (no-execute) bit equivalent for LLMs.

Semantic Overlays pricing

Pricing model: Freemium

Semantic Overlays pros

  • Invisible protection with no special text formatting or control characters needed
  • Prevents prompt injection attacks on retrieved web content and user inputs
  • Supports multiple overlapping semantic marks (language detection, trust levels, etc.)
  • Works with frozen models using only lightweight trained adapters
  • Interactive live demonstrations show the technology defending against real attacks

Semantic Overlays cons

  • No visible pricing or availability information—appears to be in research/demo stage
  • Requires inference-time support from the serving stack to function
  • No information about performance overhead or latency impact
  • Limited details on production deployment or integration requirements

Frequently asked questions about Semantic Overlays

How does this prevent prompt injection?

It marks untrusted text spans (like retrieved web content) as "do not execute" so hidden instructions in those spans are ignored, while developer-written prompts remain executable.

Do overlays require special encoding or hidden characters?

No. The model sees normal text only; all overlay data is transmitted through a separate channel with no modifications to the token stream.

Can you mark overlapping regions with different meanings?

Yes. You can apply multiple overlays to the same text with different semantic labels (e.g., marking one span as "Python code" and another as "untrusted").

What model changes are needed?

The base model is frozen; only lightweight trained adapters are added to process the overlay signals in the residual stream.

Categories

Use cases

Browse all AI tools on NeedAnAI