# PromptGuard -- Full Documentation

> AI security firewall for LLM applications. Scans every request and response for prompt injection, data leaks, jailbreaks, and PII exposure.

PromptGuard sits between your application and LLM providers, scanning every request and response through a multi-layer detection pipeline (pattern matching, ML ensemble, LLM-based analysis, policy evaluation) with sub-200ms P95 latency.

## Quick Start

- Sign up: https://app.promptguard.co
- Documentation: https://docs.promptguard.co
- API Reference: https://docs.promptguard.co/api-reference/introduction
- Python SDK: pip install promptguard-sdk
- Node.js SDK: npm install promptguard-sdk
- CLI: brew install promptguard/tap/promptguard

## Integration Methods

### Auto-Instrumentation (Recommended)

```python
import promptguard
promptguard.init(api_key="pg_sk_prod_xxx")
# All LLM calls via openai, anthropic, google-generativeai, cohere, boto3 are now secured
```

Parameters:
- api_key: PromptGuard API key (falls back to PROMPTGUARD_API_KEY env var)
- base_url: API base URL (falls back to PROMPTGUARD_BASE_URL, then https://api.promptguard.co/api/v1)
- mode: "enforce" (blocks threats) or "monitor" (logs only). Default: "enforce"
- fail_open: If True (default), requests proceed when Guard API is unreachable
- scan_responses: If True, also scan LLM responses. Default: False
- timeout: HTTP timeout in seconds. Default: 10.0

### HTTP Proxy (Drop-in)

Point your LLM SDK base URL to https://api.promptguard.co/api/v1
Pass your PromptGuard key via X-API-Key header. Your provider key stays in Authorization.

### Guard API (Standalone Scan)

POST /api/v1/guard with messages array, direction (input/output), and optional retrieved_context and media fields.

## API Endpoints

- POST /api/v1/guard -- Scan messages for threats (input/output)
- POST /api/v1/security/scan -- Standalone content scanning
- POST /api/v1/security/redact -- PII redaction
- POST /api/v1/chat/completions -- Proxied OpenAI-compatible completions
- GET /api/v1/models -- List available models

## Detection Capabilities

PromptGuard detects threats across these categories:

- Prompt Injection: Direct instruction override, role manipulation, authority impersonation, mode tricks
- Jailbreaks: 7-category LLM-based taxonomy (character obfuscation, competing objectives, lexical injection, semantic obfuscation, context obfuscation, structure obfuscation, multi-turn escalation)
- PII Detection: 39+ entity types across 10+ countries with checksum validation
- Secret Key Detection: API keys, tokens, credentials with entropy analysis
- Data Exfiltration: System prompt extraction, training data extraction, file system access
- Toxicity: Harmful content classification
- Content Safety: Multi-turn intent drift detection
- Tool Injection: Agentic workflow protection
- AI Agent Traps: 21 vectors from DeepMind's framework (HTML/CSS obfuscation, syntactic masking, image/audio steganography, font injection, framing bias, critic evasion, persona drift, RAG poisoning, memory poisoning, few-shot poisoning, sub-agent spawning, compositional fragments, sybil attacks, systemic cascade, tacit collusion, dynamic cloaking, approval fatigue)

## Security Policies

- Presets: default, support_bot, code_assistant, rag_system, data_analysis, creative_writing
- Custom rules: Natural-language business rules, topic filtering, entity blocklists
- Policy-as-code: YAML policy files for version-controlled security configuration

## Plans

- Free: 10,000 requests/month, 1 project, 1 API key, all detectors
- Pro ($99/mo): 100,000 requests/month, 5 projects, custom policies
- Scale ($199/mo): 1,000,000 requests/month, unlimited projects, analytics
- Enterprise: Custom limits, SSO, RBAC, audit logs, compliance controls

## Links

- Website: https://promptguard.co
- Dashboard: https://app.promptguard.co
- Docs: https://docs.promptguard.co
- Status: https://status.promptguard.co
- Python SDK: https://docs.promptguard.co/guides/python-sdk
- Node.js SDK: https://docs.promptguard.co/guides/node-sdk
- CLI: https://docs.promptguard.co/tools/cli
- MCP Server: https://docs.promptguard.co/tools/mcp
- API Reference: https://docs.promptguard.co/api-reference/introduction
- Support: support@promptguard.co