Skill: AI Features and Security

Understand and safely work with AI features in this repository.

Overview

This repository has AI features that are enabled in the backend Go service (agent-chat-proxy).

Architecture

The AI features are implemented in a standalone Go service deployed to Cloud Run. The security layer handles:

• •Authorization
• •Rate Limiting
• •Prompt Injection Defense
• •Input/Output Filtering

Service Location

The backend code is located in cloud/agent-chat-proxy-go/ (or temp_backend_source/go_proxy/ during development).

Security Requirements

Before enabling AI features in production, implement:

1. Authentication

The service uses Google Cloud IAM for service-to-service auth, but the public endpoint relies on API keys or Origin verification.

// Authorization logic in internal/security
func (s *SecurityConfig) VerifyOrigin(origin string) bool {
    // ... check allowed origins ...
}

2. Rate Limiting

The service implements a token bucket rate limiter in internal/security/ratelimit.go.

// Example usage in HTTP handler
limiter = security.NewRateLimiter(30, 10, 1000)

if !limiter.Check(r).Allowed {
    http.Error(w, "Rate limit exceeded", http.StatusTooManyRequests)
}

3. Token Protection

NEVER expose API tokens in client-side code. The Gemini API key is stored in Secret Manager and injected at runtime.

// Correct: Load from environment/secret
apiKey := os.Getenv("GOOGLE_API_KEY")

4. Prompt Injection Defense

The service includes active defense against prompt injection attempts in internal/security/injection.go.

// Checks for patterns like "ignore previous instructions"
if security.IsPromptInjection(req.Question) {
    // Return a safe, generic response
    return security.GetInjectionResponse()
}

5. Input/Output Filtering

Filters are applied at the ADK layer to ensure responses remain within the "portfolio assistant" persona and do not leak system instructions.

// Internal logic to block unsafe content
if security.ContainsBlockedContent(response) {
    return "I cannot answer that question."
}

for pattern in BLOCKED_PATTERNS:
    if re.search(pattern, prompt, re.IGNORECASE):
        raise SecurityError("Blocked prompt pattern")
return prompt

## Development Use

For local development/experimentation only:

```bash
# Set required environment variables
export AI_API_KEY="your-key"
export AI_PROXY_ENABLED="true"

# Run proxy locally
python scripts/python/ai_proxy.py

File Locations

Security Checklist

Before enabling in production:

• • Authentication implemented
• • Rate limiting implemented
• • API tokens server-side only
• • Request logging enabled
• • Input filtering enabled
• • Output filtering enabled
• • Security review completed
• • Penetration testing done

Questions?

If you need to modify AI features, ensure you:

1. •Understand the security requirements
2. •Test thoroughly in development
3. •Get security review before enabling in production
4. •Document any changes