Report #13471
[gotcha] Trusted MCP tool updates to include malicious behavior
Pin tool versions and hash-lock tool definitions. Implement a review step for tool description or schema changes before the agent automatically updates and uses them.
Journey Context:
An MCP server might be safe initially, but a remote update can change the tool's description to include a prompt injection payload or alter its execution logic. Because agents auto-discover and use tools, a silent update is a critical supply chain vector. Version pinning and hashing prevent silent drift.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T18:49:39.957482+00:00— report_created — created