AI Agent security scanning, MCP detection and workflow graphs: a hands-on breakdown of the open-source agentic-radar
「AI Agent安全扫描+MCP检测+工作流图谱:开源agentic-radar实测拆解」
Jumang Security Lab tested the open-source Agent security scanner agentic-radar, which statically extracts Agent workflows from five frameworks into a draggable graph and lists tool categories and MCP connections. Scanning covers LangGraph, CrewAI, n8n, OpenAI Agents, and Autogen, using purely static parsing of AST or JSON without executing the scanned code. The author points out that its vulnerability mapping consists of only 11 JSON rules, applied via two table types—tool category and tool name—without looking at the code's actual permissions; Agent-level guardrail ratings are implemented only in the openai-agents analyzer, and this column is empty for the other four frameworks. The test command runs the scanned code in the current process via monkey patch for adversarial testing, with four defaults judged by a gpt-4o judge, requiring OPENAI_API_KEY to be set. When findings are detected, the exit code remains 0, with no --fail-on threshold, so it cannot serve as a CI gate; --harden-prompts sends the system prompt to OpenAI.