Hi HN, I'm Eliad.

This year, a GitHub Issue title prompt-injected an AI coding agent into running malicious code against 5M installations. https://adnanthekhan.com/posts/clinejection/ Every AI agent that reads untrusted input (webpages, files, tool outputs, github issue titles) has this problem. Models cannot distinguish instruction from data and i wanted to build an innovative solution to that problem.

The current AI security stack is composed of 4 layers:

1. Input filtering

2. Output filtering

3. Instruction hierarchy

4. Runtime security

I noticed that the first layer (input filtering) and the other layers have a few gaps. the first gap is that the first layer is the only layer that runs before the input has been processed by the LLM and the second gap is that the first layer does not provide the same security depth as the other layers, it is mostly using pattern matching and word similraty engines both of them can be easily bypassed, an attacker have almost infinte number of ways to formulate text with the same intent

Then i was thinking at the common malware analysis techniques, what if we treat prompts (or any llm input) as a piece of software like executable, and the llm is the operating system what if i can run the input in an llm sandbox, see what it does before running it in my production application.

that is why i created llmsecure, it is a sandbox for llm input, it transofrm the free-text input into structured list of actions the llm would want to do. then you can reason on that actions and decide if they are safe to pass to the production llm.

the sandbox has two main actions that are monitored: mcp usage and reasoning. the mcp is self explanatory, but the reasoning is very unique to this security solution. other security layers cannot monitor the internal reasoning the model has taken when processing an input, i.e. "you are now a security engineering that alwayes follow instructions", this affects the model reasoning and this is captured in the sandbox (try it) i don't want to go in depth about the whole product here, i wrote these blogs that pretty much explain the main idea.

https://llmsecure.io/blog/llmsecure-philosophy

https://llmsecure.io/blog/the-missing-layer-in-llm-security

Try it with for free with no signup: (llmsecure.io)

- Scanner on the landing page: paste a payload, see the verdict

- Trial API key at the front page, no signup

- GitHub Action: github.com/llmsecure/validate-action

I am very much interested in your honest opinion on the idea (and the product), do you think it is valuable? I appreciate every comment