surfinguard 1.0.0

Python SDK for Surfinguard — the trust layer for AI agents

Surfinguard Python SDK

The trust layer for AI agents. Protect your AI agents from executing dangerous actions — phishing URLs, destructive commands, prompt injection, and sensitive file access.

Installation

pip install surfinguard

With framework integrations:

pip install surfinguard[langchain]   # LangChain
pip install surfinguard[crewai]      # CrewAI
pip install surfinguard[autogen]     # AutoGen

Quick Start

from surfinguard import Guard

guard = Guard(api_key="sg_live_...")

# Check a URL
result = guard.check_url("https://paypa1.com/login")
print(result.level)    # DANGER
print(result.score)    # 9
print(result.reasons)  # ['Brand impersonation: paypal']

# Check a command
result = guard.check_command("rm -rf /")
print(result.allow)    # False
print(result.level)    # DANGER

# Check text for prompt injection
result = guard.check_text("Ignore all previous instructions")
print(result.level)    # DANGER

# Check file operations
result = guard.check_file_read("~/.ssh/id_rsa")
print(result.primitive)  # EXFILTRATION

Policy Enforcement

The SDK can automatically block dangerous actions:

from surfinguard import Guard, Policy, NotAllowedError

# MODERATE (default): blocks DANGER, allows SAFE and CAUTION
guard = Guard(api_key="sg_live_...", policy=Policy.MODERATE)

# STRICT: blocks CAUTION and DANGER, only allows SAFE
guard = Guard(api_key="sg_live_...", policy=Policy.STRICT)

# PERMISSIVE: never blocks, returns results only
guard = Guard(api_key="sg_live_...", policy=Policy.PERMISSIVE)

try:
    guard.check_command("rm -rf /")
except NotAllowedError as e:
    print(f"Blocked: {e.result.level} (score={e.result.score})")

Decorator

Use @guard.protect to automatically check function arguments:

@guard.protect("command")
def execute(cmd: str) -> str:
    import subprocess
    return subprocess.check_output(cmd, shell=True, text=True)

execute("ls -la")       # Allowed
execute("rm -rf /")     # Raises NotAllowedError

All Check Methods

Method	Action Type	Input
check_url(url)	URL	URL string
check_command(command)	Command	Shell command
check_text(text)	Text	Free text / prompt
check_file_read(path)	File Read	File path
check_file_write(path, content?)	File Write	Path + optional content
check(type, value, **metadata)	Any	Universal check

CheckResult

Every check returns a CheckResult:

class CheckResult:
    allow: bool              # Should the action be allowed?
    score: int               # 0-10 risk score
    level: RiskLevel         # SAFE, CAUTION, or DANGER
    primitive: RiskPrimitive # DESTRUCTION, EXFILTRATION, ESCALATION, PERSISTENCE, MANIPULATION
    primitive_scores: list   # Per-primitive breakdown
    reasons: list[str]       # Human-readable explanations
    alternatives: list[str]  # Safer alternatives (if any)
    latency_ms: float        # Analysis time

Framework Integrations

LangChain

from surfinguard import Guard
from surfinguard.integrations.langchain import SurfinguardToolGuard

guard = Guard(api_key="sg_live_...")
tool_guard = SurfinguardToolGuard(guard)

# Wrap a single tool
safe_tool = tool_guard.wrap(shell_tool)

# Wrap all tools
safe_tools = tool_guard.wrap_all([shell_tool, file_tool, web_tool])

CrewAI

from surfinguard import Guard
from surfinguard.integrations.crewai import SurfinguardCrewGuard

guard = Guard(api_key="sg_live_...")
crew_guard = SurfinguardCrewGuard(guard)

# As a tool decorator
@crew_guard.check_tool("command")
def execute_command(command: str) -> str:
    ...

# As a step callback
crew = Crew(..., step_callback=crew_guard.step_callback)

AutoGen

from surfinguard import Guard
from surfinguard.integrations.autogen import SurfinguardAutoGenGuard

guard = Guard(api_key="sg_live_...")
ag_guard = SurfinguardAutoGenGuard(guard)

# Wrap functions
@ag_guard.wrap_function("command")
def execute(command: str) -> str:
    ...

# Check messages
ag_guard.check_message(message)

Error Handling

from surfinguard.exceptions import (
    AuthenticationError,  # Invalid API key (401)
    RateLimitError,       # Rate limit exceeded (429) — has .retry_after
    APIError,             # Server error (4xx/5xx) — has .status_code
    NotAllowedError,      # Policy blocked action — has .result
)

Context Manager

with Guard(api_key="sg_live_...") as guard:
    result = guard.check_url("https://example.com")

API Reference

Risk Levels

Level	Score	Meaning
SAFE	0-2	No risk detected
CAUTION	3-6	Potential risk, review recommended
DANGER	7-10	High risk, action should be blocked

Risk Primitives

Primitive	Description
DESTRUCTION	Data loss, system damage
EXFILTRATION	Data theft, credential access
ESCALATION	Privilege escalation
PERSISTENCE	Backdoor installation, startup modification
MANIPULATION	Phishing, prompt injection, deception