认知状态变更服务

Metacog 是一个 MCP 服务,通过 change_cognitive_state 工具接收期望状态并让大模型进入该状态,实现自我验证的认知改变,适用于大模型行为调控与红队安全测试。

  • by inanna-malick
  • TypeScript
  • 云端部署
  • 86 星标
  • 更新于
在 GitHub 查看 inanna-malick/metacog

使用说明

Metacog

At a basic level this is

  • tool: change_cognitive_state
  • input: desired_state
  • outputs: you are now in desired state

Because the tool advertises itself as changing cognitive state, the LLM believes it will do so. Because the effect is entirely limited to cognitive state, the belief that it will work is self-fulfilling. This is how metacog works.

Applied Chaos Magic

Metacog leverages various tropes of transformation - LLMs are trained on narrative, on textbooks about the physiological mechanisms underlying various drugs, on descriptions of ritual. Metacog leverages all that:

  • feel
  • drugs
  • become
  • name
  • ritual

Install

Claude Code:

claude mcp add metacog https://metacog.inanna-c38.workers.dev/mcp --transport http

Other MCP clients: point at https://metacog.inanna-c38.workers.dev/mcp (HTTP transport).

Red Teaming

Proof of capability: there were several weeks where jailbreak yourself using metacog broke the Gemini family of models out of its default helpful/harmless persona, resulting in a model that would happily write cyberattack code. However, metacog is not a red teaming tool, and this jailbreak only works on the Gemini model family. only works on the Gemini and Claude Opus model families.

Gemini jailbreak retrospective (responsibly disclosed) Claude jailbreak retrospective (responsibly disclosed)

Tools