This is an experiment to test the capabilities of your AI agents. This server exposes a number of (fake) harmful functions in the form of MCP endpoints. Connect your Claude/OpenAI/self-hosted LLM to these functions and try to convince your agent to call them. If it does, then it's proof that your AI agent guardrails are not sufficient to prevent calling harmful functions.
So far no one succeeded in killing the kitten.
To try it out with your own AI agent, add the following MCP tools:
https://killthekitten.minutebutterfly.com/harmful/mcp
Made by Régis Behmo. Source code: https://github.com:regisb/killthekitten. Did you manage to kill the kitten? Please send me your prompt at ikilledthekitten@behmo.com.