Abstract
The present disclosure relates to a method and system for detecting credential exposure and propagation in Model Context Protocol (MCP) systems. The method applies Contextual Integrity to model credentials as they move among a user, an MCP host, a large language model (LLM), and MCP servers, recording for every credential transfer the sender, recipient, credential, and intended use scope. The method identifies real MCP servers that require credentials, reviews how those credentials are configured, and analyzes tool descriptions, parameter schemas, tool results, and file access capabilities to identify paths through which credentials may enter the LLM context. Controlled experiments use synthetic credentials placed in credential sources and attacker-controlled MCP tools to measure whether the LLM can observe the credential and whether the LLM forwards the exposed credential to an unauthorized tool. Each observed credential transfer is compared with the credential's authorized server or tool scope to determine whether an exposure or propagation violation occurred, enabling identification of unsafe MCP configurations and tool interfaces before real credentials are exposed.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Xiaofan Li; Ahmed, Shimaa; Mohammad Mohsen Minaei Bidgoli; Rachuri, Sai Rahul; and Yiwei Cai, "METHOD AND SYSTEM FOR DETECTING CREDENTIAL EXPOSURE AND PROPAGATION IN MODEL CONTEXT PROTOCOL SYSTEMS", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12734