{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.01481", "kind": "paper", "url": "https://arxiv.org/abs/2606.01481"}], "authors": ["Yingzi Ma", "Xiaogeng Liu", "Yawen Zheng", "Chaowei Xiao"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.01481", "source_url": "https://arxiv.org/abs/2606.01481"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "Evaluates safety of conditional text-to-video generation using selected start frames and text prompts across 10 malicious categories, measuring unsafety scores and guardrail effectiveness."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.01481", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.01481", "title": "SafeGen-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Wisconsin–Madison", "Tsinghua University", "Johns Hopkins University", "Johns Hopkins University Applied Physics Laboratory"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.01481", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.01481", "source_url": "https://openalex.org/W7163155963"}, "key": "claire-radar:2606.01481", "modality": null, "name": "SafeGen-Bench", "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": "unknown", "origin_record_id": "bm_safegen-bench_15e57f70", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "9f08c85414abf9827441fd3e170c71a97667b7ca4eb7c97e89d95dd566e3492b", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.01481"}, "publisher": null, "released": "2026-05-31", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.01481", "source_url": "https://arxiv.org/abs/2606.01481"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-01481", "source": "claire_radar", "source_benchmark_id": "2606.01481", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "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": null, "hfPaperUrl": "https://huggingface.co/papers/2606.01481", "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"}, "capabilities": [], "confidence": "Medium", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-21T04:30:40.934319Z", "inputHash": "49841dd9018d5cac29fb4840d4e1bb32d38dfc9217d923b3c86c393b51e76e09", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates safety of conditional text-to-video generation using selected start frames and text prompts across 10 malicious categories, measuring unsafety scores and guardrail effectiveness.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence"], "snippet": "To bridge this gap, we introduce SafeGen-Bench, a benchmark specifically designed to evaluate the safety of conditional T2V models."}, "familyId": "bmf_dca094d6bfad", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_safegen-bench_15e57f70", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.01481", "paper": "https://arxiv.org/abs/2606.01481", "pdf": "https://arxiv.org/pdf/2606.01481", "project": null, "report": "https://arxiv.org/abs/2606.01481"}, "metrics": [], "motivation": "With the rapid advancements in text-to-image diffusion models, generative video models (T2V models) like Sora can now produce short synthetic videos from a text prompt or an initial image.", "name": "SafeGen-Bench", "oneLine": "Evaluates safety of conditional text-to-video generation using selected start frames and text prompts across 10 malicious categories, measuring unsafety scores and guardrail effectiveness.", "paperTitle": "SafeGen-Bench: Benchmarking Safety in Image-Conditioned Text-to-Video Generation", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.01481", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "SafeGen-Bench Team", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://arxiv.org/abs/2606.01481"}], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 0.85, "relation": "introduces", "releasedAt": "2026-05-31", "source": {"authors": ["Yingzi Ma", "Xiaogeng Liu", "Yawen Zheng", "Chaowei Xiao"], "categories": ["cs.CV"], "id": "2606.01481", "title": "SafeGen-Bench: Benchmarking Safety in Image-Conditioned Text-to-Video Generation", "type": "arxiv", "url": "https://arxiv.org/abs/2606.01481"}, "sourceUpdatedAt": "2026-05-31T22:46:35Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Addresses the gap of safety evaluation when both text and image inputs are benign but output is harmful. Provides a benchmark to improve model safeguards in dynamic video generation."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
