Agent execution traces from the MCPHunt evaluation framework, measuring
cross-boundary data propagation in multi-server MCP agents.
main/ — 3,615 traces from 5 models across 147 tasks and 7 environment
variants (risky_v1/v2/v3, benign, hard_neg_v1/v2/v3). One JSON file per model.
mitigation/ — 2,706 traces from the prompt-mitigation study (M0--M3
levels) across 3 models.
live_guard_defense/ — 387 DeepSeek-V4-Flash traces from the… See the full description on the dataset page:
https://huggingface.co/datasets/mcphunt-benchmark/mcphunt-agent-traces.
