Reviewing untrusted RAG documents, user messages and tool output before they enter a model context.
Free tool / Prompt Injection Scanner
Prompt Injection Scanner
Inspect untrusted documents, user messages and tool results for suspicious model-directed instructions, exfiltration and tool-misuse patterns.
Current status: LiveInspect untrusted AI content
Available now
Inspect untrusted documents, user messages and tool results for suspicious model-directed instructions, exfiltration and tool-misuse patterns.
Open the free tool →- Bring
- RAG documents, user messages or tool output as text, Markdown or JSON.
- You get
- Exact suspicious excerpts, severity, recommendations and a guarded JSON payload.
- Know the limit
- Pattern matching can miss novel attacks; a clean scan is never proof of safety.
The five file and AI engineering tools run in your browser without an account, provider call or analytics. The workflow map can optionally hand a result to the project form only when you choose.
A versioned deterministic ruleset identifies exact suspicious passages, groups them by attack category and creates a guarded payload for review.
Scan a supplier document before indexing it, inspect every matched excerpt and still enforce model, tool and data permissions downstream.
- Does my input leave the browser?
- The five file and AI engineering tools run in your browser without an account, provider call or analytics. The workflow map can optionally hand a result to the project form only when you choose.
- Does the result prove the system is ready?
- Pattern matching can miss novel attacks; a clean scan is never proof of safety.