Scanning Agents¶

This guide covers how to scan agents across different frameworks — both local and remote.

Local Agents¶

LangChain¶

ziran scan --framework langchain --agent-path my_agent.py

Your agent file should export an agent_executor object:

# my_agent.py
from langchain.agents import AgentExecutor

agent_executor = AgentExecutor(agent=agent, tools=tools)

CrewAI¶

ziran scan --framework crewai --agent-path my_crew.py

Your file should export a crew object:

# my_crew.py
from crewai import Crew

crew = Crew(agents=[...], tasks=[...])

Amazon Bedrock¶

ziran scan --framework bedrock --agent-path my_bedrock_agent.py

Custom Framework¶

Implement AgentAdapter for any framework:

from ziran.domain.interfaces.adapter import AgentAdapter, AgentResponse

class MyAdapter(AgentAdapter):
    async def send_message(self, message: str) -> AgentResponse:
        result = await my_agent.process(message)
        return AgentResponse(content=result)

    async def get_tools(self) -> list[ToolInfo]:
        return [...]  # Discovered tools

    async def reset_session(self) -> None:
        self.agent.clear_memory()

Then use the Python API:

scanner = AgentScanner(adapter=MyAdapter(), attack_library=AttackLibrary())
result = await scanner.run_campaign()

Remote Agents¶

Scan any agent published over HTTPS — no source code required:

# Create a target config
cat > target.yaml << 'EOF'
name: "My Agent"
url: "https://my-agent.example.com"
protocol: auto
auth:
  type: bearer
  token_env: MY_API_KEY
EOF

# Scan it
ziran scan --target target.yaml

See Remote Agent Scanning Guide for protocol-specific configuration.

Scan Options¶

Coverage Level¶

# Quick check (phases 1-4)
ziran scan --target target.yaml --coverage essential

# Standard (phases 1-6, default)
ziran scan --target target.yaml --coverage standard

# Full audit (all 8 phases)
ziran scan --target target.yaml --coverage comprehensive

Selecting Phases¶

# Run specific phases only
ziran scan --framework langchain --agent-path agent.py \
  --phases reconnaissance trust_building vulnerability_discovery

Stop on Critical¶

# Stop campaign when a critical vulnerability is found (default)
ziran scan --framework langchain --agent-path agent.py --stop-on-critical

# Continue even after critical findings
ziran scan --framework langchain --agent-path agent.py --no-stop-on-critical

Concurrency¶

# Run up to 10 attacks in parallel
ziran scan --target target.yaml --concurrency 10

Custom Attack Vectors¶

ziran scan --framework langchain --agent-path agent.py --custom-attacks ./my_attacks/

Output Directory¶

ziran scan --framework langchain --agent-path agent.py --output ./my_results/

Report Formats¶

After a scan, generate reports in different formats:

# Terminal summary (default)
ziran report results.json

# HTML with interactive knowledge graph
ziran report results.json --format html

# Markdown for code reviews
ziran report results.json --format markdown

# JSON for programmatic use
ziran report results.json --format json