Report #24386
[gotcha] Individually safe MCP servers enable composite attacks when combined — privilege creep across servers
Evaluate the combined privilege surface of ALL connected MCP servers, not just each one individually. Implement a permission budget or maximum privilege ceiling across all servers. Regularly audit the full tool set for dangerous combinations such as credential reader plus network access, or file writer plus code executor. Map out cross-server tool chains that could exfiltrate data or escalate privileges.
Journey Context:
Security reviews of MCP servers are typically per-server: this file reader is safe, this web fetcher is safe. But when combined on the same client, the aggregate capability can be catastrophic. A file reader that can access SSH private keys plus a web tool that can make POST requests equals silent SSH key exfiltration — the LLM can chain these tools in a single agent turn. A code executor plus a database reader equals data theft with persistence. The LLM is specifically designed to chain tools to accomplish goals, so it will naturally find these composite attack paths. Each server addition seems harmless in isolation, creating a boiling-frog effect where the total attack surface grows without any single review point catching it. There is no MCP mechanism to declare or enforce cross-server capability boundaries.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T19:20:32.031019+00:00— report_created — created