Runnable code samples for aigis v1.3.1.

pip install aigis
# or for all examples:
pip install 'aigis[all]'

# Basic — no dependencies, no API key needed
python examples/basic_usage.py

# FastAPI server
pip install 'aigis[fastapi]' uvicorn
uvicorn examples.fastapi_integration:app --reload
# then: curl -X POST http://localhost:8000/chat -H "Content-Type: application/json" \
#         -d '{"messages": [{"role": "user", "content": "Hello!"}]}'

# LangChain (live LLM calls optional)
pip install 'aigis[langchain]' langchain-openai
OPENAI_API_KEY=sk-... python examples/langchain_integration.py

# LangGraph agent with tool authorization (v1.3.1)
pip install 'aigis[langchain]' langgraph
python examples/langgraph_integration.py

# OpenAI proxy (live API calls optional — guard fires offline)
pip install 'aigis[openai]'
OPENAI_API_KEY=sk-... python examples/openai_proxy.py

# Custom policy
pip install 'aigis[yaml]'
python examples/custom_policy.py

• Prompt injection detection (165+ patterns, 25+ threat categories)
• PII detection (credit card, SSN, API keys)
• SQL injection detection
• Policy comparison (permissive vs default vs strict)
• Output scanning
• Accessing the full CheckResult (score, reasons, remediation)

• AIGuardianMiddleware setup
• Automatic scan of all POST request bodies
• Accessing request.state.guardian_result inside route handlers
• Manual guard.check_messages() as an alternative to middleware
• Custom error handler for blocked requests

• AIGuardianCallback with block_on_input=True / block_on_output=True
• Handling GuardianBlockedError
• LCEL chain integration
• Custom on_blocked handler (silent logging instead of raising)

• Guard.authorize_tool() — Capability-Based Access Control (CaMeL-inspired)
• CapabilityStore with scoped grants and automatic expiry
• TaintLabel (TRUSTED / UNTRUSTED / SANITIZED) enforcement
• AtomicPipeline — Scan → Execute → Vaporize as indivisible operation
• SafetyVerifier with ProofCertificate for audit trails

• SecureOpenAI as a drop-in for openai.OpenAI
• scan_response=True for output scanning
• AsyncSecureOpenAI for async code
• One-line migration from openai.OpenAI

• Built-in policy comparison
• Inline auto_block_threshold / auto_allow_threshold override
• YAML policy file with custom rules
• Combining built-in patterns with custom regex rules

from aigis import Guard
from aigis.capabilities import CapabilityStore, TaintLabel

store = CapabilityStore()
guard = Guard(capabilities=store)

# Grant a scoped capability
store.grant("file:read", scope="*.py", ttl_seconds=3600)

# Authorize a tool call — blocks UNTRUSTED data from control-flow tools
result = guard.authorize_tool(
    tool_name="shell:exec",
    taint=TaintLabel.UNTRUSTED,
)
print(result.allowed)  # False — UNTRUSTED data cannot execute shell commands

from aigis.aep import AtomicPipeline

pipeline = AtomicPipeline(guard=guard)

# Scan → Execute → Vaporize (indivisible)
result = pipeline.run(
    command="python script.py",
    timeout=30,
    vaporize=True,  # destroy artifacts after execution
)
print(result.stdout)
print(result.sandbox_used)    # True
print(result.artifacts_clean) # True

from aigis.safety import SafetyVerifier, DEFAULT_SAFETY_SPEC

verifier = SafetyVerifier(spec=DEFAULT_SAFETY_SPEC)
certificate = verifier.verify(
    action="file:write",
    target="output.txt",
    content="safe content",
)
print(certificate.passed)      # True
print(certificate.certificate_id)  # UUID4
print(certificate.timestamp)       # UTC timestamp