agentsentinel-crewai

Security scanner for CrewAI multi-agent workflows.
Powered by AgentSentinel on SingularityNET mainnet.

Detects 8 attack categories before your agents run:

• Agent Impersonation & Identity Spoofing
• Tool Call Hijacking (shell, exec, eval tools)
• Memory Poisoning
• Orchestrator Manipulation
• Privilege Escalation via Agent Chaining
• Goal Hijacking / Objective Corruption
• Inter-Agent Communication Injection
• Infinite Loop / Resource Exhaustion

Install

pip install agentsentinel-crewai

Quick Start

from crewai import Agent, Task, Crew
from agentsentinel_crewai import scan_crew, SecurityAudit, secure_crew

# Build your crew as normal
researcher = Agent(
    role="Researcher",
    goal="Research topics",
    backstory="Expert researcher",
    allow_delegation=True,
    tools=["web_search", "read_file"]
)

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

# Option 1 — scan and print report
report = scan_crew(crew)
report.print()

# Option 2 — raise SecurityError on CRITICAL findings
SecurityAudit(crew, block_on="CRITICAL").scan()

# Option 3 — decorator on your crew factory
@secure_crew(block_on="CRITICAL")
def build_research_crew():
    return Crew(agents=[...], tasks=[...])

crew = build_research_crew()  # raises SecurityError if CRITICAL issues found

Configuration

# Custom AgentSentinel endpoint (default: localhost:7009)
report = scan_crew(crew, endpoint="your-server:7009")

# Or via environment variable
import os
os.environ["AGENTSENTINEL_ENDPOINT"] = "your-server:7009"

Using the SingularityNET hosted service:
Pay per scan in ASI tokens at 0.005 ASI/scan via the marketplace.
→ AgentSentinel on SingularityNET

Report Output

AgentSentinel Security Report — CrewAI
============================================================
  Risk Level : CRITICAL
  Risk Score : 75/100
  Findings   : 4
  Duration   : 12ms
  Framework  : crewai

  Findings:
  [CRITICAL] AGT02 — Shell execution tool detected
             Agent: AdminAgent
             Fix: Whitelist allowed tool calls. Never allow shell/exec tools without human approval.
  [HIGH    ] AGT05 — Unrestricted delegation enabled
             Fix: Implement least-privilege per agent role.
  [HIGH    ] AGT08 — No iteration limit — infinite loop risk
             Fix: Set explicit timeouts and iteration limits.
  [MEDIUM  ] AGT01 — Verbose mode exposes agent internals
             Fix: Disable verbose in production.
============================================================

CI/CD Integration

# GitHub Actions example
- name: Security scan CrewAI workflow
  run: |
    pip install agentsentinel-crewai
    python -c "
    from agentsentinel_crewai import scan_crew
    from your_module import build_crew
    report = scan_crew(build_crew())
    if report.risk_level == 'CRITICAL':
        exit(1)
    "

Links

• NeuralSentinel Platform
• AgentSentinel on SingularityNET
• EU AI Act Compliance
• Contact: medhasni760@gmail.com

MIT License — © 2026 NeuralSentinel