MCPTox
Capture Summary
Benchmark for tool poisoning attacks on real-world MCP servers. Search result reports 45 live MCP servers, 353 tools, 1312 malicious test cases, 20 LLM agent settings, and widespread vulnerability, with low refusal rates.
Relevance
- High-value benchmark for MCP-specific agent security.
- Supports research on metadata validation, model refusal gaps, and tool-use authorization.
- Important complement to WASP and agent prompt injection benchmarks.
Collection Notes
Collected as current benchmark source.