AI app safety
The failure modes that only exist once your app calls a model.
Overview
An app that calls an LLM inherits a threat model none of the usual scanners were built for. lumioguard traces every prompt-assembly path, tool definition, and model-output sink in your repo, then flags where untrusted text reaches the model and where model output reaches something that executes it.
What it looks for
- Request input concatenated into a system or instruction prompt
- Fetched pages, database rows, or uploaded files entering prompt or tool context
- Model output executed as SQL, shell, or markup without parsing
- Tool-using agents with no scoping and no human in the loop
- Secrets or unminimised PII sent along in the prompt
Why it matters
Prompt injection stops being a party trick the moment the model holds tools. An instruction smuggled in through a fetched page can make a tool-using agent read your database, send mail, or spend money with your app’s credentials. Model-authored SQL dropped into a query runs at whatever privilege the connection carries.
Rules in this check
How the fixes arrive
Run them all on your app
Connect your repo and your live services with read-only scopes. The first scan is free, and nothing changes without your approval.