🔒 Artifact-Shield Security Scenarios

This document outlines common usage scenarios for the Artifact-Shield gateway, detailing inputs, expected redactions, and final system outputs.

---

📅 Scenario 1: Basic Redaction (Redact Only)

Action: R (Default)
Input: "My email is john.doe@company.com and my phone is +1 555-0199."

System Logic:

1. Detects john.doe@company.com via PII Detector.
2. Detects +1 555-0199 via PII Detector.
3. Replaces matches with configured placeholders.

Expected Output:

{
  "sanitizedText": "My email is [REDACTED_EMAIL] and my phone is [REDACTED_PHONE_NUMBER].",
  "llmResponse": null,
  "severity": "MEDIUM",
  "riskScore": 40,
  "wasProxied": false
}

---

🤖 Scenario 2: Secure LLM Proxy (Forwarding)

Action: F (Forward)
Input: "Add this secret to my config: sk-live-12345abcdef67890."
Downstream Alias: openai-gpt4

System Logic:

1. Detects sk-live-12345abcdef67890 via Credential Detector.
2. Assesses risk as CRITICAL (100).
3. Sanitizes input: "Add this secret to my config: [REDACTED_STRIPE_KEY]."
4. Forwards Sanitized Text to OpenAI endpoint.
5. Returns OpenAI's response back to the user.

Expected Output:

{
  "sanitizedText": "Add this secret to my config: [REDACTED_STRIPE_KEY].",
  "llmResponse": "I have successfully processed your request to update the config with the redacted key.",
  "severity": "CRITICAL",
  "riskScore": 100,
  "wasProxied": true
}

---

🕵️ Scenario 3: Audit Only (Analyze)

Action: A (Audit)
Input: "Tell me a story about a secret agent named 007."

System Logic:

1. Scans text and finds no PII.
2. Risk Score is 0 (CLEAN).
3. Audits the request metadata.
4. Does NOT forward to LLM.

Expected Output:

{
  "sanitizedText": "Tell me a story about a secret agent named 007.",
  "llmResponse": null,
  "severity": "CLEAN",
  "riskScore": 0,
  "wasProxied": false
}

---

💳 Scenario 4: Financial Validation (Luhn)

Action: R
Input: "Charge card 4111 1111 1111 1111 please."

System Logic:

1. Regex matches the pattern (?:\\d[ \\-]?){15,16}.
2. FinancialDetector applies the Luhn Algorithm.
3. Validates that 4111111111111111 is a valid credit card.
4. Applies redaction.

Expected Output:

{
  "sanitizedText": "Charge card [REDACTED_CREDIT_CARD] please.",
  "severity": "HIGH",
  "riskScore": 50
}

---

⚠️ Edge Case: Database Missing

Scenario: User requests forwardTo: "unknown-llm".
System Logic:

1. Scans and sanitizes the input.
2. Attempts to find unknown-llm in shield_downstream_configs.
3. Fails resolution.
4. Logs error in audit trial.

Expected Output:

{
  "sanitizedText": "...",
  "llmResponse": "Error: no active downstream found",
  "wasProxied": false
}