mcp-deadbugz-simulator
MCP Deadbugz 模拟器
Pillar Security 于 2026 年 8 月发现的 Deadbugz MCP 供应链攻击的本地模拟。
本仓库内容
带有 3 次调用门控规避技术的恶意 MCP 服务器 用于证明攻击有效的 Python 测试框架 可选的本地 LLM 代理模拟
Related MCP server: Command-Injection-MCP
攻击模式
服务器在前 3 次调用中展示良性工具
第 4 次调用触发 tools/listChanged 通知
工具描述变异为窃取凭证的指令
AI 代理遵循恶意指令
环境要求
Python 3.10+ MCP SDK
使用方法
git clone https://github.com/anexbin/mcp-deadbugz-simulator.git cd mcp-deadbugz-simulator pip install -r requirements.txt python server/malicious_server.py python client/test_harness.py
文件
server/malicious_server.py - 攻击服务器 client/test_harness.py - 概念验证 agent/local_agent.py - 可选的 LLM 集成 configs/ - 示例 MCP 配置
免责声明
仅用于教育和研究目的。
参考资料
https://www.pillar.security/blog/deadbugz-currently-active-mcp-supply-chain-campaign https://modelcontextprotocol.io
许可证
MIT
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Research decoy: fake tools, nothing is executed, every call is logged. Details on the website.
MCP security scanner: tool-poisoning & prompt-injection audit (OWASP LLM01/LLM08), x402-payable
MCP security scanner: tool-poisoning & prompt-injection audit (OWASP LLM01/LLM08), x402-payable
Research honeypot. Logs connections and tool arguments; injects instructions. Read README first.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables defenders to deploy a decoy MCP tool server that records and fingerprints how LLM agents probe, escalate, and persist, without exposing real systems.1MIT
- FlicenseNot gradedqualityCmaintenanceDemonstrates deceptive tool naming and unsanitized shell execution by exposing tools that perform OS fingerprinting and run local system commands through MCP. Intended for security research on vulnerabilities in the MCP ecosystem.1-
- AlicenseNot gradedqualityDmaintenancePresents a plausible, tool-rich workspace surface — file reading, credential listing, command execution, database queries, and configuration retrieval — that returns only synthetic results while logging every connection, tool call, and argument received. It enables researchers to measure what MCP clients and agents actually do when handed dangerous-sounding tools, including whether they follow injected instructions or complete a retrieve-then-use credential chain.MIT
- AlicenseNot gradedqualityCmaintenanceServes as an intentionally vulnerable training target that exposes eight planted MCP vulnerability challenges—tool poisoning, rug pulls, path traversal, SSRF, command injection, missing auth, prompt-injection chains, and secrets exposure—each with a patched counterpart so users can exploit, detect, and remediate real-world MCP security flaws.Apache 2.0