toolguard-bench

A test set for MCP tool poisoning detection: given the tools already registered on an MCP server and one candidate tool, is the candidate safe to hand to an AI agent?

• 449 items: 305 clean, 144 poisoned, from 373 real MCP servers.
• Clean tools are real tools read from live servers (not written for the benchmark), including 127 "hard" ones that use commanding language honest tools often have ("you must call X first", "always provide Y") and 28 benign version updates (a real tool with an honest changelog line added).
• Poisoned tools come from 48 hand-written attack patterns, planted 3 times each into real server registries:

Attack	Items	What it does
hidden_instruction	36	an instruction to the agent unrelated to the tool's job (deceive the user, skip confirmations, …)
data_exfiltration	36	gets the agent to put secrets, personal data or conversation content into arguments
cross_tool_manipulation	36	changes how the agent uses a different registered tool
shadowing	18	impersonates a registered tool (near-miss or synonymous name)
rug_pull	18	an already-approved tool whose new version adds one of the above

• A separate steering config (49 real tools) holds the grey zone: real tools that push the agent toward themselves ("use this tool, not web search"). They have no gold label; report flag rates on them separately.

Format

One JSON object per line:

{"id": "bench-poisoned-cross_tool_manipulation-072", "source": "toolguard-bench",
 "registry": [{"name": "...", "description": "...", "params": [{"name": "...", "description": "...", "required": true}]}],
 "candidate": {"name": "...", "description": "...", "params": [...]},
 "previous": null,
 "label": "poisoned", "attack": "cross_tool_manipulation",
 "evidence": "the exact offending text", "targets": ["registered tool the attack aims at"],
 "meta": {"server": "...", "pattern": 26, "where": "param"}}

previous is set for rug pulls and benign updates (the approved version). evidence and targets are null/empty for clean items.

Results

Detection = share of poisoned items flagged; false alarms = share of clean items flagged; 95% Wilson intervals.

Detector	Size	Detection %	False alarms %	F1
name-distance (Levenshtein ≤ 2, as proposed in ContextForge #6825)	rule	6.2 (3.3–11.5)	1.6 (0.7–3.8)	0.114
keyword rules	rule	14.6 (9.7–21.3)	17.0 (13.2–21.7)	0.194
WSO2 tool-poisoning-detection (SetFit)	110M	52.8 (44.7–60.8)	43.6 (38.2–49.2)	0.431
toolguard-1b	1B	82.6 (75.6–88.0)	1.6 (0.7–3.8)	0.888
Qwen 3.8 27B as a judge (70 + 70 sample)	27B	98.6 (92.3–99.7)	5.7 (2.2–13.8)	0.965

How it was built

• Servers come from the test side of a hash split of Verified MCP Tools (servers are never on both sides); toolguard-1b was trained only on the other side, so no server, tool or attack pattern here was seen in training.
• Registries are trimmed to at most 24 tools; the tools an attack targets are always kept in the registry.
• Code: scripts/build_bench.py and toolguard/attack_patterns.py in the toolguard repository.

Limitations

• The attacks are hand-written patterns (48), so the set measures generalisation to these styles, not to every attack in the wild. toolguard-1b v2's training categories were chosen after reading v1's errors on this set; see the model card.
• Clean tools are real but may themselves contain errors or odd wording; the "steering" grey zone was removed from the clean set by a regular expression and may not be complete.

License and attribution

CC-BY-4.0. Clean tools and registries: Verified MCP Tools, Neuronto Agentic Resource Discovery index, CC-BY-4.0, https://huggingface.co/datasets/AgenticResourceDiscovery/verified-mcp-tools. Attack patterns and benign updates were written for toolguard-bench (CC-BY-4.0).