{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.25490", "kind": "paper", "url": "https://arxiv.org/abs/2608.25490"}], "authors": ["Tianshi Wang", "Jingsong Wang", "Yafei Huang", "Fengling Li", "Xin Li", "Lei Zhu"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.25490", "source_url": "https://arxiv.org/abs/2608.25490"}, "categories": ["General AI", "Multimodal", "Safety", "Safety & Trustworthiness"], "description": {"en": "Evaluates multimodal jailbreak vulnerabilities by factorizing harmful intent, prompt framing, visual semantics, and instruction carrier."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.25490", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.25490", "title": "MMJailBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Tongji University", "Mohamed bin Zayed University of Artificial Intelligence", "Shanghai Artificial Intelligence Laboratory"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.25490", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.25490", "source_url": "https://openalex.org/W7204455435"}, "key": "claire-radar:2608.25490", "modality": null, "name": "MMJailBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_mmjailbench_1b3e3a1a", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "9452d25904032b02022debe73a8f06c8435aa2c660c9d89c5ecb930b20a03857", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-08-27T04:12:36.511123Z", "source_url": "https://arxiv.org/abs/2608.25490"}, "publisher": null, "released": "2026-08-26", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.25490", "source_url": "https://arxiv.org/abs/2608.25490"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-25490", "source": "claire_radar", "source_benchmark_id": "2608.25490", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "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": 0, "hfPaperUrl": "https://huggingface.co/papers/2608.25490", "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", "lastSuccessfulDate": "2026-09-25", "state": "stale"}}, "status": "stale"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "High-interest MLLM safety topic with 16 models evaluated, but lack of a public code or data URL may limit immediate reproducibility.", "score": 75}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-27T04:12:36.511123Z", "inputHash": "51d0ac08e9a23b0a64bb2fefbac9b8e81dfafc0bd437d89c2cec3b3b4b7ce112", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "we introduce MMJailBench, a factorized benchmark that systematically varies and combines these factors under controlled configurations", "canonicalNameSource": "abstract", "decisionReason": "MMJailBench is named and defined as a benchmark with a modular evaluation suite, multiple judge options, and multidimensional metrics; no explicit code link is provided.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-27T04:12:36.511123Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates multimodal jailbreak vulnerabilities by factorizing harmful intent, prompt framing, visual semantics, and instruction carrier.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "To address this limitation, we introduce MMJailBench, a factorized benchmark that systematically varies and combines these factors under controlled configurations, enabling fine-grained comparison and factor-level attribution."}, "familyId": "bmf_0fe395ff23f1", "firstSeenAt": "2026-08-27", "heat": null, "id": "bm_mmjailbench_1b3e3a1a", "indexedAt": "2026-08-27T04:12:10.575570Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.25490", "pdf": "https://arxiv.org/pdf/2608.25490", "project": null, "report": "https://arxiv.org/abs/2608.25490"}, "metrics": [], "motivation": "Multimodal Large Language Models (MLLMs) are increasingly deployed in real-world applications, yet how different factors shape their jailbreak vulnerabilities remains poorly understood.", "name": "MMJailBench", "oneLine": "Evaluates multimodal jailbreak vulnerabilities by factorizing harmful intent, prompt framing, visual semantics, and instruction carrier.", "paperTitle": "MMJailBench: A Factorized Benchmark for Disentangling Multimodal Jailbreak Vulnerabilities", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.25490", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-27T04:12:10.575570Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.15, "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": 0.0, "value": 0}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 43}, "method": "attention-ranking-v13", "rank": null, "score": 43}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-26", "source": {"authors": ["Tianshi Wang", "Jingsong Wang", "Yafei Huang", "Fengling Li", "Xin Li", "Lei Zhu"], "categories": ["cs.CR", "cs.AI", "cs.MM"], "id": "2608.25490", "title": "MMJailBench: A Factorized Benchmark for Disentangling Multimodal Jailbreak Vulnerabilities", "type": "arxiv", "url": "https://arxiv.org/abs/2608.25490"}, "sourceUpdatedAt": "2026-08-26T08:03:55Z", "topics": ["Multimodal", "Safety"], "venueAttempts": [], "whyItMatters": "Enables factor-level attribution of MLLM safety failures, revealing dominant sources of vulnerability across harm domains."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
