Hey Umberto,
Went through rag-chatbot — the part where retrieved Markdown chunks go into the LLM context before the response is the classic indirect injection surface. A single poisoned .md file in the corpus can instruct the model before your system prompt stops it.
I built Zentric Protocol for exactly this: one POST before each LLM call, 23ms, CLEARED/BLOCKED verdict. Deterministic — same input, same result, no model drift.
100k free requests to test on your actual corpus:
https://zentricprotocol.com/quickstart
Worth a look?
Abel
Hey Umberto,
Went through rag-chatbot — the part where retrieved Markdown chunks go into the LLM context before the response is the classic indirect injection surface. A single poisoned .md file in the corpus can instruct the model before your system prompt stops it.
I built Zentric Protocol for exactly this: one POST before each LLM call, 23ms, CLEARED/BLOCKED verdict. Deterministic — same input, same result, no model drift.
100k free requests to test on your actual corpus:
https://zentricprotocol.com/quickstart
Worth a look?
Abel