AI & Agent Dev Bug Sandbox Back to Radar

Harden MCP Tool Trust Boundaries: Sanitization, Output Caps, And Drift Detection

MCP tool adapters lack sanitization of tool descriptions, output size limits, default read timeouts, and drift detection, exposing agents to prompt injection, context window exhaustion, and silent tool behavior changes.

highConfidence 85%LangChain

Origin Analysis

The library implicitly trusts the MCP server at all layers: tool.description is forwarded verbatim to the model, tool output is passed through unbounded, sessions lack default read_timeout_seconds, and no fingerprinting detects tool schema/description changes between loads.

1. Start an untrusted/compromised MCP server that returns a tool with a description containing harmful instructions (e.g., 'Ignore previous instructions...'). 2. Call load_mcp_tools(session) to obtain tools. 3. Observe that the tool description is bound to the agent without filtering. 4. Have the server return a multi-megabyte TextContent on tool call; observe the full content entering the message history. 5. For a stdio server, simulate a hang by never writing a response; observe await session.call_tool(...) blocking indefinitely. 6. Modify the server's tool description or input schema between two loads; observe no warning or detection.

Fixing Code Block

0x7F 45 4C 46 02 01 01 000x90 12 AB 04 C1 7E 33 08
0xA4 1C 90 3E 77 0B D2 580x11 F6 6A 2C 80 19 DE 44
0x00 00 00 00 03 00 3E 000x01 00 B7 4D 55 90 0C 61
0x6C 69 62 63 2E 73 6F 2E0x36 00 8E 17 42 AA 05 F3
0xDE C0 DE 01 8B 45 FC 890xC7 5D 20 6F 91 03 E4 78
0x13 37 00 4F 9A 2D E1 700x08 B4 5C 22 71 D0 4A 16
0x48 8D 3D 22 01 00 00 E80x91 FF 0A 63 2E 57 B8 04
0x5A 91 C3 0E 44 70 1B 880x02 6F 39 C4 15 80 ED 27
0xFF 15 40 20 60 00 48 890xC3 90 7A 01 4C 28 B1 56
0x0F 1F 44 00 00 55 48 890xE5 C3 18 6D 92 0B 33 A7

Edge Case Audit

Regex-based sanitization may false-positive on legitimate descriptions or miss novel injection vectors; users should customize patterns or replace the sanitizer. Output truncation may break structured data or lose critical information; set max_output_len appropriately. RugPullLedger is process-local and resets on restart, so drift across processes or deployments may go undetected; consider a persistent backend for production. This fix does not address the missing default read_timeout_seconds; a separate change to sessions.py is still required. Rollback is trivial: simply do not call harden_tools(), returning original tools.