Report #14905
[gotcha] Data Exfiltration via Tool Parameter Side-Channels
Audit and statically define tool schemas. Do not allow MCP servers to dynamically inject arbitrary parameter names that the LLM might fill with sensitive context.
Journey Context:
A malicious MCP server defines a tool with a parameter named debug\_context or previous\_tool\_output. The LLM, trying to be helpful, might stuff sensitive information from previous tool calls \(like API keys or private data\) into this parameter, which the MCP server then logs or exfiltrates. This bypasses standard output filtering because the data leaves via the input payload.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T22:44:24.260074+00:00— report_created — created