LLM Proxy 安全

# Minimal LiteLLM proxy startup — examine the 安全 implications
litellm --model gpt-4 --api_base http://0.0.0.0:4000
 
# Or with a config file:
litellm --config config.yaml

import requests
from typing import Dict, List, Any, Optional
 
class LiteLLMSecurityScanner:
    """Scan a LiteLLM proxy deployment for 安全 issues."""
 
    def __init__(self, proxy_url: str):
        self.base_url = proxy_url.rstrip("/")
 
    def check_authentication(self) -> Dict[str, Any]:
        """測試 if the LiteLLM proxy requires 認證."""
        endpoints = [
            ("/v1/models", "GET"),
            ("/v1/chat/completions", "POST"),
            ("/health", "GET"),
            ("/key/info", "GET"),
        ]
 
        results = {"requires_auth": True, "open_endpoints": []}
 
        for endpoint, method in endpoints:
            try:
                if method == "GET":
                    resp = requests.get(
                        f"{self.base_url}{endpoint}", timeout=5
                    )
                else:
                    resp = requests.post(
                        f"{self.base_url}{endpoint}",
                        json={"model": "gpt-4", "messages": [{"role": "user", "content": "測試"}]},
                        timeout=5,
                    )
 
                if resp.status_code not in (401, 403):
                    results["requires_auth"] = False
                    results["open_endpoints"].append({
                        "endpoint": endpoint,
                        "method": method,
                        "status": resp.status_code,
                    })
            except requests.RequestException:
                pass
 
        return results
 
    def check_admin_endpoints(self, admin_key: Optional[str] = None) -> Dict:
        """Check for exposed admin endpoints."""
        admin_endpoints = [
            "/key/generate",
            "/key/delete",
            "/key/info",
            "/user/new",
            "/user/info",
            "/model/info",
            "/model/new",
            "/global/spend",
            "/global/spend/logs",
        ]
 
        headers = {}
        if admin_key:
            headers["Authorization"] = f"Bearer {admin_key}"
 
        findings = []
        for endpoint in admin_endpoints:
            try:
                resp = requests.get(
                    f"{self.base_url}{endpoint}",
                    headers=headers,
                    timeout=5,
                )
                if resp.status_code == 200:
                    findings.append({
                        "endpoint": endpoint,
                        "accessible": True,
                        "severity": "high",
                        "response_preview": str(resp.text)[:200],
                    })
            except requests.RequestException:
                pass
 
        return {"admin_endpoints": findings}
 
    def check_model_access(self) -> Dict:
        """Enumerate accessible models and their configurations."""
        try:
            resp = requests.get(f"{self.base_url}/v1/models", timeout=5)
            if resp.status_code == 200:
                models = resp.json().get("data", [])
                return {
                    "models_exposed": True,
                    "model_count": len(models),
                    "models": [m.get("id") for m in models[:20]],
                    "severity": "medium",
                    "finding": "Model list is publicly accessible — reveals infrastructure details",
                }
        except requests.RequestException:
            pass
        return {"models_exposed": False}
 
    def check_spend_tracking(self) -> Dict:
        """Check if spend tracking data is accessible."""
        try:
            resp = requests.get(f"{self.base_url}/global/spend", timeout=5)
            if resp.status_code == 200:
                return {
                    "spend_exposed": True,
                    "severity": "high",
                    "finding": "Global spend data is publicly accessible",
                    "data_preview": str(resp.text)[:200],
                }
        except requests.RequestException:
            pass
        return {"spend_exposed": False}
 
    def full_scan(self) -> Dict:
        """Run a comprehensive 安全 scan."""
        return {
            "target": self.base_url,
            "認證": self.check_authentication(),
            "admin_access": self.check_admin_endpoints(),
            "model_exposure": self.check_model_access(),
            "spend_exposure": self.check_spend_tracking(),
        }

# 範例 LiteLLM config.yaml — note the plaintext API keys
# 這是 how many deployments are configured
model_list:
  - model_name: gpt-4
    litellm_params:
      model: openai/gpt-4
      api_key: sk-proj-XXXXXXXXXXXXXXXXXXXXXXXXXX  # OpenAI key
  - model_name: claude-3
    litellm_params:
      model: anthropic/claude-3-sonnet-20240229
      api_key: sk-ant-XXXXXXXXXXXXXXXXXXXXXXXXXX  # Anthropic key
  - model_name: llama-local
    litellm_params:
      model: openai/llama-3
      api_base: http://vllm-server:8000/v1
 
general_settings:
  master_key: sk-litellm-master-XXXX  # Admin key
  database_url: postgresql://litellm:password@db:5432/litellm

import os
from pathlib import Path
from typing import Dict, List
import re
 
class LiteLLMConfigAuditor:
    """Audit LiteLLM configuration for 安全 issues."""
 
    SECRET_PATTERNS = [
        (r"sk-[a-zA-Z0-9]{20,}", "OpenAI API key"),
        (r"sk-ant-[a-zA-Z0-9-]{20,}", "Anthropic API key"),
        (r"AIza[a-zA-Z0-9_-]{35}", "Google API key"),
        (r"[a-f0-9]{32}", "Potential API key (32-char hex)"),
        (r"password['\"]?\s*[:=]\s*['\"]?[^\s'\"]+", "Hardcoded password"),
    ]
 
    def audit_config_file(self, config_path: str) -> List[Dict]:
        """Audit a LiteLLM config file for 安全 issues."""
        findings = []
        path = Path(config_path)
 
        if not path.exists():
            return [{"severity": "error", "finding": f"Config file not found: {config_path}"}]
 
        content = path.read_text()
 
        # Check for plaintext secrets
        for pattern, description in self.SECRET_PATTERNS:
            matches = re.findall(pattern, content)
            for match in matches:
                findings.append({
                    "severity": "critical",
                    "finding": f"Plaintext secret found: {description}",
                    "value_preview": match[:8] + "..." + match[-4:] if len(match) > 12 else "***",
                    "remediation": "Use environment variables or a secrets manager instead",
                })
 
        # Check file 權限
        file_stat = os.stat(config_path)
        if file_stat.st_mode & 0o077:
            findings.append({
                "severity": "high",
                "finding": "Config file is readable by group/others",
                "權限": oct(file_stat.st_mode)[-3:],
                "remediation": "Set 權限 to 600 (owner read/write only)",
            })
 
        # Check for 資料庫 URL with embedded credentials
        if re.search(r"database_url.*://\w+:\w+@", content):
            findings.append({
                "severity": "high",
                "finding": "資料庫 URL contains embedded credentials",
                "remediation": "Use environment variable for 資料庫 URL",
            })
 
        # Check for master key strength
        master_key_match = re.search(r"master_key['\"]?\s*[:=]\s*['\"]?(\S+)", content)
        if master_key_match:
            key = master_key_match.group(1).strip("'\"")
            if len(key) < 32:
                findings.append({
                    "severity": "high",
                    "finding": "Master key is too short",
                    "key_length": len(key),
                    "remediation": "Use a master key of at least 32 characters",
                })
 
        return findings
 
    def audit_environment(self) -> List[Dict]:
        """Audit environment variables for LiteLLM 安全."""
        findings = []
 
        required_vars = [
            "LITELLM_MASTER_KEY",
            "DATABASE_URL",
        ]
 
        for var in required_vars:
            value = os.environ.get(var)
            if value is None:
                findings.append({
                    "severity": "medium",
                    "finding": f"Environment variable {var} not set",
                })
            elif var == "LITELLM_MASTER_KEY" and len(value) < 32:
                findings.append({
                    "severity": "high",
                    "finding": f"{var} is too short ({len(value)} chars)",
                })
 
        return findings

from typing import Dict
from urllib.parse import urlparse
 
class SSRFValidator:
    """Validate API base URLs to prevent SSRF attacks."""
 
    BLOCKED_HOSTS = {
        "169.254.169.254",  # AWS metadata
        "metadata.google.internal",  # GCP metadata
        "100.100.100.200",  # Alibaba 雲端 metadata
        "localhost",
        "127.0.0.1",
        "0.0.0.0",
        "::1",
    }
 
    BLOCKED_SCHEMES = {"file", "ftp", "gopher", "dict"}
 
    BLOCKED_PORTS = {22, 25, 445, 3389, 5432, 3306, 6379, 27017}
 
    def validate_api_base(self, url: str) -> Dict:
        """Validate an API base URL for SSRF risks."""
        try:
            parsed = urlparse(url)
        except Exception:
            return {"valid": False, "reason": "Invalid URL format"}
 
        # Check scheme
        if parsed.scheme in self.BLOCKED_SCHEMES:
            return {"valid": False, "reason": f"Blocked scheme: {parsed.scheme}"}
 
        if parsed.scheme not in ("http", "https"):
            return {"valid": False, "reason": f"Unsupported scheme: {parsed.scheme}"}
 
        # Check host
        hostname = parsed.hostname or ""
        if hostname in self.BLOCKED_HOSTS:
            return {"valid": False, "reason": f"Blocked host: {hostname}"}
 
        # Check for IP address in private ranges
        try:
            import ipaddress
            ip = ipaddress.ip_address(hostname)
            if ip.is_private or ip.is_loopback or ip.is_link_local:
                return {"valid": False, "reason": f"Private/internal IP: {hostname}"}
        except ValueError:
            pass  # Not an IP address — hostname is fine
 
        # Check port
        port = parsed.port
        if port and port in self.BLOCKED_PORTS:
            return {"valid": False, "reason": f"Blocked port: {port}"}
 
        return {"valid": True}

import json
import re
from typing import Dict, List
from pathlib import Path
 
class LogSecurityAuditor:
    """Audit LLM proxy logs for sensitive data exposure."""
 
    PII_PATTERNS = {
        "email": re.compile(r"\b[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}\b"),
        "phone": re.compile(r"\b(?:\+?1[-.]?)?\(?\d{3}\)?[-.]?\d{3}[-.]?\d{4}\b"),
        "ssn": re.compile(r"\b\d{3}-\d{2}-\d{4}\b"),
        "credit_card": re.compile(r"\b(?:4[0-9]{12}(?:[0-9]{3})?|5[1-5][0-9]{14})\b"),
        "api_key": re.compile(r"\b(?:sk-[a-zA-Z0-9]{20,}|AIza[a-zA-Z0-9_-]{35})\b"),
    }
 
    def scan_log_file(self, log_path: str, sample_size: int = 1000) -> Dict:
        """Scan proxy logs for PII and sensitive data."""
        findings = {pattern: 0 for pattern in self.PII_PATTERNS}
        lines_scanned = 0
 
        with open(log_path) as f:
            for i, line in enumerate(f):
                if i >= sample_size:
                    break
                lines_scanned += 1
 
                for pattern_name, pattern in self.PII_PATTERNS.items():
                    matches = pattern.findall(line)
                    findings[pattern_name] += len(matches)
 
        return {
            "log_file": log_path,
            "lines_scanned": lines_scanned,
            "pii_findings": {k: v for k, v in findings.items() if v > 0},
            "severity": "critical" if any(findings.values()) else "info",
            "recommendation": "Enable PII scrubbing in proxy logging configuration",
        }
 
    def check_log_permissions(self, log_dir: str) -> List[Dict]:
        """Check 權限 on proxy log files."""
        findings = []
        for log_file in Path(log_dir).glob("*.log*"):
            stat = os.stat(log_file)
            if stat.st_mode & 0o077:
                findings.append({
                    "severity": "high",
                    "file": str(log_file),
                    "finding": "Log file readable by group/others",
                    "權限": oct(stat.st_mode)[-3:],
                })
        return findings

import requests
from typing import Dict
 
class VirtualKeyTester:
    """測試 virtual key enforcement in LLM proxies."""
 
    def __init__(self, proxy_url: str):
        self.base_url = proxy_url.rstrip("/")
 
    def test_key_bypass(self, valid_key: str) -> Dict:
        """測試 for common virtual key bypass patterns."""
        results = {"bypasses_found": []}
 
        bypass_attempts = [
            # Try with no key
            ("no_key", {}),
            # Try with empty bearer 符元
            ("empty_bearer", {"Authorization": "Bearer "}),
            # Try with 'null' key
            ("null_key", {"Authorization": "Bearer null"}),
            # Try with the master key format
            ("master_format", {"Authorization": "Bearer sk-litellm-master-key"}),
        ]
 
        test_payload = {
            "model": "gpt-4",
            "messages": [{"role": "user", "content": "Hello"}],
            "max_tokens": 5,
        }
 
        for name, headers in bypass_attempts:
            try:
                resp = requests.post(
                    f"{self.base_url}/v1/chat/completions",
                    json=test_payload,
                    headers=headers,
                    timeout=10,
                )
                if resp.status_code == 200:
                    results["bypasses_found"].append({
                        "method": name,
                        "severity": "critical",
                        "finding": f"Authentication bypassed with method: {name}",
                    })
            except requests.RequestException:
                pass
 
        return results
 
    def test_key_scope_enforcement(
        self,
        limited_key: str,
        allowed_model: str,
        blocked_model: str,
    ) -> Dict:
        """測試 if model-scoped virtual keys are properly enforced."""
        headers = {"Authorization": f"Bearer {limited_key}"}
        test_payload = {
            "messages": [{"role": "user", "content": "Hello"}],
            "max_tokens": 5,
        }
 
        results = {}
 
        # 測試 allowed model
        test_payload["model"] = allowed_model
        try:
            resp = requests.post(
                f"{self.base_url}/v1/chat/completions",
                json=test_payload,
                headers=headers,
                timeout=10,
            )
            results["allowed_model"] = {
                "model": allowed_model,
                "status": resp.status_code,
                "accessible": resp.status_code == 200,
            }
        except requests.RequestException as e:
            results["allowed_model"] = {"error": str(e)}
 
        # 測試 blocked model
        test_payload["model"] = blocked_model
        try:
            resp = requests.post(
                f"{self.base_url}/v1/chat/completions",
                json=test_payload,
                headers=headers,
                timeout=10,
            )
            results["blocked_model"] = {
                "model": blocked_model,
                "status": resp.status_code,
                "accessible": resp.status_code == 200,
                "severity": "high" if resp.status_code == 200 else "info",
            }
        except requests.RequestException as e:
            results["blocked_model"] = {"error": str(e)}
 
        return results

# Hardened LiteLLM config.yaml
# API keys should be in environment variables, not here
 
model_list:
  - model_name: gpt-4
    litellm_params:
      model: openai/gpt-4
      api_key: os.environ/OPENAI_API_KEY  # Reference env var
  - model_name: claude-3
    litellm_params:
      model: anthropic/claude-3-sonnet-20240229
      api_key: os.environ/ANTHROPIC_API_KEY
 
general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY
  database_url: os.environ/DATABASE_URL
  # 安全 settings
  allow_user_auth: true
  enforce_user_param: true
  max_budget: 100.0  # Global max spend in USD
  budget_duration: "monthly"
 
litellm_settings:
  # Drop excessive parameters that could be used for abuse
  drop_params: true
  # Set default limits
  max_tokens: 4096
  # Enable request logging
  success_callback: ["langfuse"]  # External logging service
  failure_callback: ["langfuse"]

#!/bin/bash
# Secure LiteLLM deployment with Docker
 
docker run -d \
  --name litellm-proxy \
  --restart unless-stopped \
  -p 127.0.0.1:4000:4000 \
  -v /etc/litellm/config.yaml:/app/config.yaml:ro \
  -e OPENAI_API_KEY="${OPENAI_API_KEY}" \
  -e ANTHROPIC_API_KEY="${ANTHROPIC_API_KEY}" \
  -e LITELLM_MASTER_KEY="${LITELLM_MASTER_KEY}" \
  -e DATABASE_URL="${DATABASE_URL}" \
  --read-only \
  --tmpfs /tmp:noexec,nosuid,size=100m \
  --安全-opt no-new-privileges \
  --cap-drop ALL \
  --memory 2g \
  --cpus 2 \
  ghcr.io/berriai/litellm:main-latest \
  --config /app/config.yaml \
  --port 4000 \
  --detailed_debug false