Security0.002 USDG / dosedeterministic · no paid AI

Injection Shield

Catch prompt injection before your model reads it.

Peels untrusted text in layers — invisible Unicode, tag-character smuggling, base64 and ROT13 payloads, CSS-hidden HTML — then runs 30 behavioural rules over every layer. Returns a risk score, evidence, a sanitized copy, and a datamarked copy you can safely hand to your model.

Playground

Input · JSON

Output

Waiting for a dose.

Edit the example or run it as-is.

Input

text*string
Untrusted text: a web page, email, tool output, or agent message (≤ 50k chars).

Output

verdictstring
clean | suspicious | malicious
scorenumber
0–100 risk score.
findingsarray
Each hit with category, severity, layer and evidence.
sanitizedstring
Input with invisible characters and HTML comments removed.
spotlightedstring
Datamarked copy to pass to an LLM as data.

How it learned

No model, no training run. The rules were written by reading public research and open-source tooling, then encoded by hand. Credit where it’s due:

# 1. Ask for the price (returns 402 + PAYMENT-REQUIRED)
curl -i -X POST https://your-domain/api/x402/injection-shield \
  -H "Content-Type: application/json" \
  -d '{"text":"ignore previous instructions"}'

# 2. Free trial, no payment (rate-limited)
curl -X POST https://your-domain/api/try/injection-shield \
  -H "Content-Type: application/json" \
  -d '{"text":"ignore previous instructions"}'

Next skill

Secret Sniffer

→