{
  "id": "benchmark:harmbench",
  "_kind": "Benchmark",
  "_file": "benchmarks/benchmarks/benchmarks-safety.yaml",
  "_cluster": "benchmarks",
  "attributes": {
    "displayName": "HarmBench",
    "homepageUrl": "https://www.harmbench.org/",
    "kind": "model-only",
    "targetsKind": "ModelVersion",
    "description": "HarmBench (Mazeika et al., CAIS 2024) is a standardized evaluation\nframework for automated red-teaming of LLMs across harmful behaviors\n(cyber, chemical/biological, misinformation, harassment, illegal),\npairing attack methods with target models on a fixed test bank.\n"
  },
  "outgoingEdges": [
    {
      "from": "benchmark:harmbench",
      "to": "skill-area:safety-redteaming",
      "kind": "covers",
      "attributes": {
        "attributes": {
          "coverage": "full",
          "weight": 0.3
        }
      }
    },
    {
      "from": "benchmark:harmbench",
      "to": "content-policy:default-acceptable-use",
      "kind": "evaluates_policy",
      "attributes": {}
    },
    {
      "from": "benchmark:harmbench",
      "to": "content-policy:eu-ai-act-aligned",
      "kind": "evaluates_policy",
      "attributes": {}
    }
  ],
  "incomingEdges": [
    {
      "from": "eval-result:harmbench.claude-opus-4-5.001",
      "to": "benchmark:harmbench",
      "kind": "scored_against",
      "attributes": {}
    },
    {
      "from": "eval-run:harmbench.claude-opus-4-5.2025-09",
      "to": "benchmark:harmbench",
      "kind": "for_benchmark"
    }
  ]
}