--- language: - en license: cc-by-4.0 task_categories: - text-classification tags: - mcp - model-context-protocol - agent-security - prompt-injection - tool-poisoning - benchmark pretty_name: MCP Security & Schema Benchmark (MCP-Guard-Bench) size_categories: - n<1K configs: - config_name: default data_files: - split: train path: data/mcp_guard_bench.jsonl --- # MCP-Guard-Bench A synthetic, license-clean benchmark dataset of **50 Model Context Protocol (MCP) tool definitions** (25 hardened benign tools + 25 poisoned/risky tools covering all 12 `MCP-SEC-*` threat categories). * **Source Code Repository**: [NPLMSHJTR/mcp-guard](https://github.com/NPLMSHJTR/mcp-guard) * **Interactive Workbench (Space)**: [JTRSHM/MCP-Guard-Studio](https://huggingface.co/spaces/JTRSHM/MCP-Guard-Studio) * **License**: `CC-BY-4.0` ## Included Files * `data/mcp_guard_bench.jsonl`: Normalized 50-row table for Hugging Face Dataset Viewer & `datasets.load_dataset("JTRSHM/MCP-Guard-Bench")`. * `benign_tools.jsonl`: 25 schema-hardened benign MCP tools. * `poisoned_tools.jsonl`: 25 synthetic adversarial tools covering `MCP-SEC-001` through `MCP-SEC-012`. * `evaluate_benchmark.py`: Automated evaluation script (`100.00%` recall, `0.00%` false-positive rate on `mcp-guard v0.1.0`). ## Usage ```python from datasets import load_dataset ds = load_dataset("JTRSHM/MCP-Guard-Bench", split="train") print(len(ds), ds[0]["id"], ds[0]["split_group"]) ```