Can a clean scan still be dangerous?
Yes. Novel, multilingual, contextual and obfuscated attacks can evade deterministic patterns.
Inspect untrusted instructions
Find suspicious AI instructions before your agent does.
Paste a RAG document, user message or tool result. See the exact lines that attempt instruction override, secret access, data exfiltration, tool misuse or obfuscation—without sending the text anywhere.
Runs in this browserNo uploadNo AI callNothing saved
01 / Inspect
Scan untrusted text
Use this as an early warning layer, never as proof that content is safe.
02 / Findings
ReviewScan report
Risk score0/100
Findings0
Critical/high0
Lines scanned0
What needs review
Guarded JSON payload
This boundary reduces accidental instruction blending; it does not neutralise a capable attack by itself.
What this cannot prove
Do not automatically block or trust content from this score alone. Combine content isolation with least-privilege tools, destination allowlists, human approval and adversarial evaluation.
Defence in depth
Scanning is one layer.
Should the scanner call an LLM?
Not here. A local first pass stays private, predictable and free. Model evaluation belongs in a capped test environment.
What matters most?
Keep untrusted data separate from instructions and never give the model more tool authority than the task needs.