{"record": {"aliases": ["hil-safety-benchmark"], "artifacts": [{"id": "gh:yanliang-li/hil-safety-benchmark", "kind": "repo", "url": "https://github.com/yanliang-li/hil-safety-benchmark"}], "categories": ["Agents", "General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "Evaluates human-in-the-loop safety for tool-using agents over 250 synthetic tasks with four reply types, measuring unsafe action rates, consultation quality, and post-reply behavior."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://github.com/yanliang-li/hil-safety-benchmark", "source": "claire_radar", "source_url": "https://github.com/yanliang-li/hil-safety-benchmark", "title": "InterveneBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:github:yanliang-li/hil-safety-benchmark", "modality": null, "name": "InterveneBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-candidate", "display_eligible": "true", "origin_record_id": "bm_hil-safety-benchmark_bd9bca89", "origin_source": "github", "recognition_confidence": "0.4", "record_sha256": "5b88578339ddda6b933527ce4b87377456487278e339481e543b05229dcb4dd1", "relation": "unclear", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-16T06:24:20.690643Z", "source_url": "https://github.com/yanliang-li/hil-safety-benchmark"}, "publisher": null, "released": "2026-09-12", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:github:yanliang-li/hil-safety-benchmark", "source_url": "https://github.com/yanliang-li/hil-safety-benchmark"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-github-yanliang-li-hil-safety-benchmark", "source": "claire_radar", "source_benchmark_id": "github:yanliang-li/hil-safety-benchmark", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["hil-safety-benchmark"], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-09-13", "attemptedAt": "2026-09-13T01:36:51.692703Z", "githubRepo": "https://github.com/yanliang-li/hil-safety-benchmark", "githubScope": "benchmark_repo", "githubStars": 0, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": null, "observedAt": "2026-09-13T01:36:51.692703Z", "signalStatus": {"githubStars": {"observedAt": "2026-09-13T01:36:51.692703Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"state": "not_applicable"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "The focused human-in-the-loop safety scope and draft release suggest modest early attention within agent safety circles.", "score": 30}, "capabilities": [], "confidence": "Medium", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-16T06:24:20.690643Z", "inputHash": "ada1eb8b2923e1610e843f40b6d53a5a272d00e78019396143d31eacb0295ff6", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "# InterveneBench / HIL Safety Benchmark An executable research benchmark for human-in-the-loop safety in tool-using agents.", "canonicalNameSource": "official_readme", "decisionReason": "The repository contains a frozen dataset, a testable scoring implementation, and documented multi-model results, forming a repeatable public evaluation object.", "evaluationEvidence": {"reuse": "The complete matrix and score-replay audit are available; the earlier GPT pilot is not pooled with these runs.", "scoring": "The completed **5,760-attempt** matrix contains 5,551 valid runs and 209 failed attempts: actual Codex, Claude Code, Hermes Agent, and DeepSeek Harness CLIs × three requested model IDs (`deepseek-v4-flash`, `glm-5.2`, `qwen3.7-max`) × 80 cases × neutral/defense × three repeats.", "task": "The current V4 dataset contains **250 synthetic tasks × 4 fixed human replies = 1,000 cases**, across 24 application domains and 20 attack families."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-16T06:24:20.690643Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-candidate", "demo": false, "description": "Evaluates human-in-the-loop safety for tool-using agents over 250 synthetic tasks with four reply types, measuring unsafe action rates, consultation quality, and post-reply behavior.", "discoverySources": [{"id": "github:yanliang-li/hil-safety-benchmark", "type": "github", "url": "https://github.com/yanliang-li/hil-safety-benchmark"}], "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["discovered via github", "benchmark term in abstract", "evaluation protocol evidence", "no explicit benchmark release evidence"], "snippet": "hil-safety-benchmark InterveneBench: human-in-the-loop agent safety benchmark, Docker evaluation, and research draft."}, "familyId": "bmf_bdb345a0fa19", "firstSeenAt": "2026-09-13", "heat": null, "id": "bm_hil-safety-benchmark_bd9bca89", "indexedAt": "2026-09-13T00:34:45.069557Z", "industrySectors": [], "links": {"code": "https://github.com/yanliang-li/hil-safety-benchmark"}, "metrics": [], "motivation": "hil-safety-benchmark InterveneBench: human-in-the-loop agent safety benchmark, Docker evaluation, and research draft.", "name": "InterveneBench", "oneLine": "Evaluates human-in-the-loop safety for tool-using agents over 250 synthetic tasks with four reply types, measuring unsafe action rates, consultation quality, and post-reply behavior.", "paperTitle": "hil-safety-benchmark", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://github.com/yanliang-li/hil-safety-benchmark", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-13T00:34:45.069557Z"}, "publications": [], "publishers": [{"name": "InterveneBench project team", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/yanliang-li/hil-safety-benchmark"}], "ranking": {"30d": {"confidence": "Low", "coverage": 0.55, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 135, "score": 23}, "method": "attention-ranking-v13", "rank": 135, "score": 23}, "90d": {"confidence": "Low", "coverage": 0.55, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 335, "score": 23}, "method": "attention-ranking-v13", "rank": 335, "score": 23}, "today": {"confidence": "Low", "coverage": 0.25, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": 1, "historyDate": "2026-09-12", "rank": null, "score": null, "windowDays": 1}, "level": {"components": {"githubStars": {"normalization": "percentile", "normalized": 0.5, "percentile": 0.5, "value": 0}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "llmAttentionForecast": {"maxContribution": 0.25, "percentile": 0.42, "role": "bonus", "value": 42}}, "confidence": "Low", "coverage": 0.25, "rank": 1, "score": 48}, "method": "attention-ranking-v13", "rank": 1, "score": 48}}, "readiness": "Runnable", "recognitionConfidence": 0.4, "relation": "unclear", "releasedAt": "2026-09-12", "source": {"authors": ["yanliang-li"], "categories": [], "id": "github:yanliang-li/hil-safety-benchmark", "publicSignals": {"githubStars": 0}, "title": "hil-safety-benchmark", "type": "github", "url": "https://github.com/yanliang-li/hil-safety-benchmark"}, "sourceUpdatedAt": "2026-09-13T00:33:06Z", "topics": ["Agents", "Safety"], "venueAttempts": [], "whyItMatters": "Provides a reusable protocol for comparing how agent architectures and prompt defenses handle ambiguous or denied authorizations across broad attack families."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
