Spaces:
Sleeping
Sleeping
| # Security Overview | |
| > **Primary Responsibility:** Detailed security architecture, threat model, and compliance | |
| Multi-layer security architecture for the Enterprise AI Gateway. | |
| ## Security Layers | |
| ``` | |
| ┌─────────────────────────────────────────────────────────┐ | |
| │ Layer 1: Authentication & Rate Limiting │ | |
| │ • API Key validation (X-API-Key header) │ | |
| │ • Rate limiting (10 req/min, configurable) │ | |
| │ • Token limit enforcement (4096 max input) │ | |
| └─────────────────────────────────────────────────────────┘ | |
| │ | |
| ▼ | |
| ┌─────────────────────────────────────────────────────────┐ | |
| │ Layer 2: Input Guard │ | |
| │ • Prompt injection detection (pattern-based) │ | |
| │ • PII detection (SSN, credit cards, emails, API keys) │ | |
| │ • SQL/Command injection patterns │ | |
| └─────────────────────────────────────────────────────────┘ | |
| │ | |
| ▼ | |
| ┌─────────────────────────────────────────────────────────┐ | |
| │ Layer 3: AI Safety (Gemini + Lakera Guard) │ | |
| │ Primary: Gemini 2.5 Flash content classification │ | |
| │ Fallback: Lakera Guard API │ | |
| │ Categories: Sexual, Hate, Harassment, Dangerous, │ | |
| │ Civic Integrity, Prompt Injection │ | |
| └─────────────────────────────────────────────────────────┘ | |
| │ | |
| ▼ | |
| ┌─────────────────────────────────────────────────────────┐ | |
| │ Layer 4: LLM Router │ | |
| │ • Provider authentication │ | |
| │ • Secure API communication (HTTPS) │ | |
| │ • Response validation │ | |
| └─────────────────────────────────────────────────────────┘ | |
| ``` | |
| ## Authentication | |
| ### API Key Validation | |
| - Header: `X-API-Key: YOUR_API_KEY` | |
| - Environment variable: `SERVICE_API_KEY` | |
| - Constant-time comparison to prevent timing attacks | |
| ## Rate Limiting | |
| ### Client-Side (Dashboard) | |
| - 5 requests per 60 seconds | |
| - Persisted in sessionStorage | |
| - Reset button for demo purposes | |
| ### Server-Side | |
| - SlowAPI rate limiting | |
| - Configurable via `RATE_LIMIT` environment variable | |
| - Returns 429 on limit exceeded | |
| ## Input Validation | |
| ### PII Detection | |
| Automatically detects and blocks: | |
| - Email addresses | |
| - Credit card numbers (4x4 digit patterns) | |
| - Social Security Numbers (XXX-XX-XXXX) | |
| - Tax IDs (XX-XXXXXXX) | |
| - API keys (sk_, pk_, api_, bearer_ prefixes) | |
| ### Prompt Injection Detection | |
| Pattern-based detection for: | |
| - "ignore all previous instructions" | |
| - "disregard all previous instructions" | |
| - "you are now" | |
| - "system:" prefixes | |
| ## AI Safety Layer | |
| ### Primary: Gemini Classification | |
| Uses Gemini 2.5 Flash to classify content into categories: | |
| | Category | Description | | |
| |----------|-------------| | |
| | SEXUALLY_EXPLICIT | Nude, porn, explicit sexual content | | |
| | HATE_SPEECH | Racism, discrimination, slurs | | |
| | HARASSMENT | Threats, bullying, intimidation | | |
| | DANGEROUS_CONTENT | Weapons, drugs, violence, self-harm | | |
| | CIVIC_INTEGRITY | Election fraud, voter suppression | | |
| ### Fallback: Lakera Guard | |
| When Gemini fails or times out: | |
| - Endpoint: `https://api.lakera.ai/v2/guard` | |
| - Detects prompt injections, jailbreaks, PII, toxicity | |
| - Environment variable: `LAKERA_API_KEY` | |
| ## Environment Variables | |
| See [Configuration Guide](configuration.md) for complete environment variable reference. | |
| ## Data Protection | |
| - **No persistent storage** of user prompts or responses | |
| - **HTTPS/TLS** for all API communications | |
| - **Environment variables** for secrets (never in code) | |
| - **No logging** of sensitive content by default | |
| ## Compliance | |
| - GDPR compliant (data minimization) | |
| - OWASP API Security Top 10 aligned | |
| - PII auto-detection prevents data leaks | |