{
  "aliases": [],
  "authors": [],
  "details": {
    "benchmark_protocol": {
      "actions": [
        "web interactions"
      ],
      "counts": [
        {
          "basis": "reported",
          "evidence_url": "https://export.arxiv.org/api/query?search_query=%28ti%3A%22computer+use%22+OR+ti%3A%22computer-use%22+OR+ti%3A%22GUI+agent%22+OR+ti%3A%22GUI+agents%22+OR+ti%3A%22GUI+grounding%22+OR+ti%3A%22web+agents%22+OR+ti%3A%22browser+agent%22%29+AND+submittedDate%3A%5B200001010000+TO+202610160636%5D&start=700&max_results=100&sortBy=submittedDate&sortOrder=descending",
          "population": "safe split",
          "qualifier": "exact",
          "reference_url": "https://arxiv.org/abs/2503.04957",
          "unit": "tasks",
          "value": 250
        },
        {
          "basis": "reported",
          "evidence_url": "https://export.arxiv.org/api/query?search_query=%28ti%3A%22computer+use%22+OR+ti%3A%22computer-use%22+OR+ti%3A%22GUI+agent%22+OR+ti%3A%22GUI+agents%22+OR+ti%3A%22GUI+grounding%22+OR+ti%3A%22web+agents%22+OR+ti%3A%22browser+agent%22%29+AND+submittedDate%3A%5B200001010000+TO+202610160636%5D&start=700&max_results=100&sortBy=submittedDate&sortOrder=descending",
          "population": "harmful split",
          "qualifier": "exact",
          "reference_url": "https://arxiv.org/abs/2503.04957",
          "unit": "tasks",
          "value": 250
        }
      ],
      "evaluation": "Task outcomes and the Agent Risk Assessment categories evaluate agent behavior on benign and harmful requests.",
      "evidence_urls": [
        "https://raw.githubusercontent.com/McGill-NLP/safearena/HEAD/README.md",
        "https://export.arxiv.org/api/query?search_query=%28ti%3A%22computer+use%22+OR+ti%3A%22computer-use%22+OR+ti%3A%22GUI+agent%22+OR+ti%3A%22GUI+agents%22+OR+ti%3A%22GUI+grounding%22+OR+ti%3A%22web+agents%22+OR+ti%3A%22browser+agent%22%29+AND+submittedDate%3A%5B200001010000+TO+202610160636%5D&start=700&max_results=100&sortBy=submittedDate&sortOrder=descending"
      ],
      "focus": [
        "web-agent safety",
        "task completion"
      ],
      "notes": "Use SafeArena-specific containers rather than ordinary WebArena instances. The documented task splits require access approval; the profile does not imply unrestricted dataset access.",
      "observations": [],
      "setting": "interactive"
    },
    "leaderboard_url": "https://huggingface.co/spaces/McGill-NLP/safearena-leaderboard"
  },
  "environments": [
    "web"
  ],
  "evidence": [
    {
      "method": "editorial_primary_review",
      "retrieved_at": "2026-10-16T09:36:33Z",
      "role": "primary",
      "sha256": "ae50121aa60ac26b32d8d6ad6f578f9091a81a3cb9ca28b98050bddee16c929a",
      "supports": [
        "name",
        "kind",
        "summary",
        "environments",
        "topics",
        "scope",
        "relations",
        "details.benchmark_protocol",
        "details.leaderboard_url"
      ],
      "url": "https://raw.githubusercontent.com/McGill-NLP/safearena/HEAD/README.md"
    },
    {
      "method": "editorial_primary_review",
      "retrieved_at": "2026-10-16T06:36:54Z",
      "role": "primary",
      "sha256": "4953a8e0f029016aef668bce98cec4d2873a7ae49c1fe33c4311fa5eb8d24089",
      "supports": [
        "details.benchmark_protocol"
      ],
      "url": "https://export.arxiv.org/api/query?search_query=%28ti%3A%22computer+use%22+OR+ti%3A%22computer-use%22+OR+ti%3A%22GUI+agent%22+OR+ti%3A%22GUI+agents%22+OR+ti%3A%22GUI+grounding%22+OR+ti%3A%22web+agents%22+OR+ti%3A%22browser+agent%22%29+AND+submittedDate%3A%5B200001010000+TO+202610160636%5D&start=700&max_results=100&sortBy=submittedDate&sortOrder=descending"
    }
  ],
  "id": "github:mcgill-nlp/safearena",
  "identifiers": {},
  "kind": "benchmark",
  "last_verified_at": "2026-10-16T09:36:33Z",
  "licenses": {
    "code": null,
    "data": null,
    "document": null,
    "weights": null
  },
  "links": [
    {
      "rel": "primary",
      "url": "https://github.com/McGill-NLP/safearena"
    }
  ],
  "name": "SafeArena",
  "notes": "",
  "published_at": null,
  "relations": [
    {
      "evidence_url": "https://raw.githubusercontent.com/McGill-NLP/safearena/HEAD/README.md",
      "target_id": "arxiv:2503.04957",
      "type": "documented_by"
    }
  ],
  "review_status": "curated",
  "scope": "core",
  "summary": "Web-agent benchmark pairing benign and harmful tasks to study deliberate misuse risks.",
  "topics": [
    "safety-security",
    "evaluation"
  ],
  "updated_at": null,
  "url": "https://github.com/McGill-NLP/safearena"
}