{"record": {"aliases": ["EvoHarmBench: Breaking Content Moderation with Iterative Human-Like Evasion"], "artifacts": [{"id": "arxiv:2608.27844", "kind": "paper", "url": "https://arxiv.org/abs/2608.27844"}], "categories": ["Factuality", "General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "Evaluates content moderation systems against iteratively evolved adversarial samples across 229 semantic sub-clusters and five violation categories, with scoring based on attack success rate under readability constraints."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.27844", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.27844", "title": "EvoHarmBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.27844", "modality": null, "name": "EvoHarmBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "false", "origin_record_id": "bm_evoharmbench-breaking-content-moderation-w_fad69c38", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "14308d23d63114142059bedb76ee0420465785feef6b861f7032766c9f3d2692", "relation": "evaluates_only", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-06T08:46:12.648160Z", "source_url": "https://arxiv.org/abs/2608.27844"}, "publisher": null, "released": "2026-08-28", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.27844", "source_url": "https://arxiv.org/abs/2608.27844"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-27844", "source": "claire_radar", "source_benchmark_id": "2608.27844", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["EvoHarmBench: Breaking Content Moderation with Iterative Human-Like Evasion"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.27844", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "The topic of dynamic adversarial evaluation for content moderation is narrow, and the absence of a listed repository or dataset limits discoverability.", "score": 35}, "capabilities": ["Factuality"], "confidence": "Medium", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-06T08:46:12.648160Z", "inputHash": "9bac5855becdf83fa7b805d80234963e659afc909277f39b7e783db2c28d233e", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "we present EvoHarmBench, the first dynamic adversarial evaluation framework for content moderation systems.", "canonicalNameSource": "abstract", "decisionReason": "No artifact links or code/data release are supplied, and the paper only states a future intention to release; a clear public reuse path is therefore not yet established.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-06T08:46:12.648160Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates content moderation systems against iteratively evolved adversarial samples across 229 semantic sub-clusters and five violation categories, with scoring based on attack success rate under readability constraints.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "evaluation protocol evidence", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "To the best of our knowledge, we present EvoHarmBench, the first dynamic adversarial evaluation framework for content moderation systems."}, "familyId": "bmf_b9d84094b186", "firstSeenAt": "2026-08-31", "heat": null, "id": "bm_evoharmbench-breaking-content-moderation-w_fad69c38", "indexedAt": "2026-08-31T01:03:30.163531Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.27844", "pdf": "https://arxiv.org/pdf/2608.27844", "project": null, "report": "https://arxiv.org/abs/2608.27844"}, "metrics": [], "motivation": "Existing evaluations of harmful content detection rely predominantly on static benchmarks, which struggle to reflect the interactive adversarial ecosystem of real-world content platforms where users continuously revise their expressions in response to moderation feedback.", "name": "EvoHarmBench", "oneLine": "Evaluates content moderation systems against iteratively evolved adversarial samples across 229 semantic sub-clusters and five violation categories, with scoring based on attack success rate under readability constraints.", "paperTitle": "EvoHarmBench: Breaking Content Moderation with Iterative Human-Like Evasion", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to the Findings of EMNLP 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2608.27844", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "Findings of EMNLP 2026", "verifiedAt": "2026-08-31T01:03:30.163531Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.0, "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": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 0.55, "relation": "evaluates_only", "releasedAt": "2026-08-28", "source": {"authors": ["Ruijie Jian", "Benlei Cui", "Ting Ma", "Haidong Ding", "Kangwei Liu", "Ziwen Xu", "Longtao Huang", "Hui Xue", "Ziqiang Zhu", "Junjie Li", "Haiwen Hong"], "categories": ["cs.CL"], "id": "2608.27844", "title": "EvoHarmBench: Breaking Content Moderation with Iterative Human-Like Evasion", "type": "arxiv", "url": "https://arxiv.org/abs/2608.27844"}, "sourceUpdatedAt": "2026-08-28T02:25:43Z", "topics": ["cs.CL"], "venueAttempts": [{"decisionRaw": "Accepted to the Findings of EMNLP 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-08-31T01:03:30.163531Z", "rawValue": "Accepted to the Findings of EMNLP 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2608.27844"}], "reviewStatus": "accepted", "venueName": "Findings of EMNLP 2026"}], "whyItMatters": "Addresses the gap between static moderation benchmarks and live adversarial evasion by testing whether moderation systems remain robust when attacks adapt over multiple optimization iterations."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
