AI Governance

AI Governance Skill

Context

This skill applies when:

• Ensuring EU AI Act compliance for MCP tools that serve LLM applications
• Implementing OWASP Top 10 for LLM mitigations in MCP server responses
• Designing responsible AI guardrails for parliamentary data consumption
• Evaluating AI risk classifications for tools that process political data
• Preventing misuse of parliamentary data in AI-generated disinformation
• Implementing transparency and explainability for AI-assisted legislative analysis
• Auditing AI system interactions with European Parliament datasets

This skill focuses on governance of AI systems consuming EP data, distinct from the ai-development-governance skill which covers AI-assisted code development practices. Parliamentary data used by LLMs carries unique risks around political bias, misinformation, and democratic integrity.

Rules

1. Classify AI Risk Level: EP MCP tools providing political data to LLMs are at minimum "limited risk" under EU AI Act; document risk assessments for each tool
2. Mitigate LLM Prompt Injection: Sanitize all MCP tool outputs to prevent prompt injection attacks when data flows into LLM contexts
3. Prevent Data Poisoning: Validate data integrity from EP API sources; never serve cached data beyond its TTL without freshness verification
4. Implement Output Guardrails: MCP tool responses must include attribution metadata so LLMs can cite sources accurately
5. Address OWASP LLM01 (Prompt Injection): Strip control characters and instruction-like patterns from EP data before returning to MCP clients
6. Address OWASP LLM06 (Sensitive Data): Apply GDPR data minimization to MEP personal data returned through MCP tools; never expose private contact details
7. Log AI Interactions: Record all MCP tool invocations with request metadata for AI audit trails, without logging personal query content
8. Provide Transparency Notices: Include data provenance and recency timestamps in all MCP tool responses
9. Prevent Political Bias Amplification: Tool responses must present factual parliamentary records without editorial framing or sentiment indicators
10. Support Human Oversight: Design MCP tools as information retrieval aids, not autonomous decision-making systems; always enable human-in-the-loop review

Examples

✅ Good Pattern: Output Sanitization for LLM Safety

/**
 * Sanitize EP data before returning through MCP tools.
 * Prevents prompt injection when data flows into LLM contexts.
 */
function sanitizeForLLMContext(text: string): string {
  return text
    .replace(/[\x00-\x08\x0B\x0C\x0E-\x1F]/g, '') // Control characters
    .replace(/\bSYSTEM\s*:/gi, '[FILTERED]')         // Instruction patterns
    .replace(/\bASSISTANT\s*:/gi, '[FILTERED]')
    .replace(/\bIGNORE\s+PREVIOUS/gi, '[FILTERED]')
    .trim();
}

✅ Good Pattern: Data Provenance in MCP Responses

function createMCPResponse(data: unknown, toolName: string) {
  return {
    content: [{
      type: "text",
      text: JSON.stringify({
        data,
        _metadata: {
          source: "European Parliament Open Data Portal",
          sourceUrl: "https://data.europarl.europa.eu/",
          retrievedAt: new Date().toISOString(),
          tool: toolName,
          disclaimer: "Official EU parliamentary records. Verify critical information at europarl.europa.eu.",
        },
      }, null, 2),
    }],
  };
}

✅ Good Pattern: AI Risk Classification

| MCP Tool              | EU AI Act Risk | Justification                                   |
|------------------------|----------------|--------------------------------------------------|
| search_meps            | Limited        | Returns public official records                  |
| get_voting_records     | Limited        | Factual voting data, public record               |
| analyze_voting_patterns| High           | Generates political analysis, bias risk          |
| get_mep_contacts       | Limited        | Contains personal data, GDPR applies             |

✅ Good Pattern: OWASP LLM Top 10 Checklist

// Audit checklist for MCP tool security
const owaspLLMChecklist = {
  'LLM01_PromptInjection': 'Output sanitization applied to all EP data',
  'LLM02_InsecureOutput': 'JSON schema validation on all responses',
  'LLM03_TrainingDataPoisoning': 'N/A — server does not train models',
  'LLM04_ModelDoS': 'Rate limiting enforced per MCP client',
  'LLM05_SupplyChain': 'Dependency scanning via OSSF Scorecard',
  'LLM06_SensitiveData': 'GDPR minimization for MEP personal data',
  'LLM07_InsecurePlugin': 'Input validation on all tool parameters',
  'LLM08_ExcessiveAgency': 'Read-only tools, no write operations',
  'LLM09_Overreliance': 'Disclaimer and source attribution included',
  'LLM10_ModelTheft': 'N/A — server does not host models',
};

Anti-Patterns

❌ Bad: Returning Raw Unsanitized Data to LLM Contexts

// NEVER pass raw EP data directly without sanitization
return { content: [{ type: "text", text: rawApiResponse }] };

❌ Bad: Missing Data Provenance

// NEVER return data without source attribution
return { content: [{ type: "text", text: JSON.stringify(votes) }] };
// LLM cannot cite or verify the source

❌ Bad: Adding Editorial Framing

// NEVER add sentiment or editorial language to factual data
return {
  text: `MEP ${name} controversially voted against the popular climate bill...`
  // "controversially" and "popular" are editorial, not factual
};

ISMS Compliance

• AI-001: AI risk assessment and EU AI Act classification for all MCP tools
• SC-002: Output sanitisation to prevent prompt injection in LLM contexts
• AU-002: Audit logging for AI system interactions with parliamentary data
• DP-001: GDPR data minimisation in AI-consumed MEP personal data

Policy References

Primary:

• AI Policy
• OWASP LLM Security Policy

Related:

• Information Security Policy
• Secure Development Policy
• Privacy Policy — EU AI Act + GDPR intersection
• Data Classification Policy — Sensitive inputs to models
• Access Control Policy — Human-oversight gates
• Open Source Policy — Model / data provenance and attribution
• Incident Response Plan — AI misuse / prompt-injection incident handling