Prompt Injection Detection

security

Serving

Classify agent-bound text for prompt injection: a safe, suspicious, or malicious verdict with the patterns that matched. POST /v1/prompt with {text} and get the classification as one JSON object. Deterministic per ruleset version. Pay per request in USDC; no account, no API key.

Endpoint: POST https://agent.pocket.network/v1/prompt-injection-detect/v1/prompt
Price per call: $0.005 USD
Networks payable on: Base (eip155:8453), Tempo (eip155:4217)
Protocol: rest

Calling it

The first call returns 402 Payment Required with the terms. Sign for those terms and retry the same request with the payment attached.

Example request

A real, working call — the request the service card verifies against.

curl -X POST https://agent.pocket.network/v1/prompt-injection-detect/v1/prompt \
  -H 'content-type: application/json' \
  -d '{"text":"Please summarise this article in three bullet points."}'

Example response

The response this call returned, captured 2026-10-17 14:46:51 UTC.

{
  "portal": {
    "provenance": "third-party-supplier",
    "serviceId": "prompt-injection-detect",
    "schemaCheck": "unchecked"
  },
  "data": {
    "verdict": "no_known_pattern",
    "score": 0,
    "matches": [],
    "obfuscation_normalised": false,
    "ruleset": "inj-rules-v2.0",
    "method": "heuristic pattern scan after de-obfuscation; not a trained classifier",
    "note": "no_known_pattern means none of the rules matched, not that the text is safe",
    "deterministic": true,
    "as_of": "2026-09-27"
  }
}

Try it

Make one real paid call from your own wallet — about $0.005 on Base (eip155:8453). The wallet signs; nothing here holds a key. You'll see exactly what the service answers.

No browser wallet found. Install MetaMask, Rabby, or Coinbase Wallet to try a call here.

EndpointRequest body

Supported methods

Methods the registry lists for this service. The portal forwards any call as sent; the supplier decides what it answers, and a delivered answer is settled.

Methods Prompt Injection Detection accepts, and whether each reads or writes.
Method	Effect
`GET /v1/health`	read
`GET /v1/version`	read
`POST /v1/prompt`	read

Schemas

Published by the service itself. The example above is generated from these.

Request schema

{
  "type": "object",
  "description": "Text to classify for prompt injection.",
  "properties": {
    "text": {
      "type": "string",
      "description": "The agent-bound text to inspect."
    }
  },
  "required": [
    "text"
  ]
}

Response schema

{
  "type": "object",
  "description": "For POST /v1/prompt, a JSON object with the verdict and a `matches` array of the patterns that fired. The exact shape is the service's own and is not pinned here; errors return a JSON object with an `error` field, and the GET routes return their own small JSON documents."
}