{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.26793", "kind": "paper", "url": "https://arxiv.org/abs/2606.26793"}, {"id": "gh:fujitsuresearch/mirror", "kind": "repo", "url": "https://github.com/FujitsuResearch/mirror"}], "categories": ["Language & Knowledge", "Robotics & Embodied AI", "cs.CR"], "description": {"en": "ART-SafeBench evaluates attacks against multimodal agentic RAG systems across four attack surfaces, using attack success rate and query cost as comparable outcomes."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.26793", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.26793", "title": "ART-SafeBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.26793", "modality": null, "name": "ART-SafeBench", "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": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_art-safebench_b8bbb590", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "6e9b5977adc7dfd9fe1fdbb8afbbbecc06302f296b9e5fae90fcd8c90004ee83", "relation": "evaluates_only", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T20:23:07.437049Z", "source_url": "https://arxiv.org/abs/2606.26793"}, "publisher": null, "released": "2026-06-25", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.26793", "source_url": "https://arxiv.org/abs/2606.26793"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-26793", "source": "claire_radar", "source_benchmark_id": "2606.26793", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Robotics & Embodied AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Agentic RAG security, cross-surface attacks, a sizable released package, and public code provide several credible paths to near-term attention.", "score": 66}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T20:23:07.437049Z", "inputHash": "d0cd379dfc9fc8a223d1ba49bc070ff37016fb786dbad15696f32c756896b0d3", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We release ART-SafeBench with 41,815 in-package records and runtime adapters yielding 41,991+ total records across four surfaces.", "canonicalNameSource": "abstract", "decisionReason": "The benchmark is formally named, released with records and runtime adapters, linked through a public repository, and evaluated with concrete ASR and query-cost outcomes.", "evaluationEvidence": {"reuse": "Code and ART-SafeBench benchmark: https://github.com/FujitsuResearch/mirror", "scoring": "MIRROR attains 76% ASR on image poisoning compared with 52% for baselines, 97% ASR on orchestrator attacks at half the query cost", "task": "Across four attack surfaces on a multimodal agentic RAG target"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T20:23:07.437049Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "ART-SafeBench evaluates attacks against multimodal agentic RAG systems across four attack surfaces, using attack success rate and query cost as comparable outcomes.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "public artifact URL", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "We release ART-SafeBench with 41,815 in-package records and runtime adapters yielding 41,991+ total records across four surfaces."}, "familyId": "bmf_396a5399116a", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_art-safebench_b8bbb590", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Robotics"], "links": {"code": "https://github.com/FujitsuResearch/mirror", "data": null, "paper": "https://arxiv.org/abs/2606.26793", "pdf": "https://arxiv.org/pdf/2606.26793", "project": null, "report": "https://arxiv.org/abs/2606.26793"}, "metrics": [], "motivation": "Multimodal agentic retrieval-augmented generation (RAG) systems expand the attack surface beyond prompt injection to include text poisoning, image injection, direct-query attacks, and orchestrator-level tool manipulation.", "name": "ART-SafeBench", "oneLine": "ART-SafeBench evaluates attacks against multimodal agentic RAG systems across four attack surfaces, using attack success rate and query cost as comparable outcomes.", "paperTitle": "MIRROR: Novelty-Constrained Memory-Guided MCTS Red-Teaming for Agentic RAG", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": "6 pages, 2 figures. Accepted at the 2026 International Joint Conference on Neural Networks (IJCNN 2026), IEEE WCCI 2026; presented as an oral talk. Code and ART-SafeBench benchmark: https://github.com/FujitsuResearch/mirror", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2606.26793", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "2026 International Joint Conference on Neural Networks (IJCNN 2026), IEEE WCCI 2026", "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "Fujitsu Research", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/FujitsuResearch/mirror"}], "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "evaluates_only", "releasedAt": "2026-06-25", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Inderjeet Singh", "Andr\\'es Murillo", "Motoyoshi Sekiya", "Yuki Unno", "Junichi Suga"], "categories": ["cs.CR", "cs.AI", "cs.LG"], "id": "2606.26793", "title": "MIRROR: Novelty-Constrained Memory-Guided MCTS Red-Teaming for Agentic RAG", "type": "arxiv", "url": "https://arxiv.org/abs/2606.26793"}, "sourceUpdatedAt": "2026-06-25T09:26:49Z", "topics": ["cs.CR"], "venueAttempts": [{"decisionRaw": "6 pages, 2 figures. Accepted at the 2026 International Joint Conference on Neural Networks (IJCNN 2026), IEEE WCCI 2026; presented as an oral talk. Code and ART-SafeBench benchmark: https://github.com/FujitsuResearch/mirror", "evidence": [{"level": "author-claim", "observedAt": "2026-08-19T11:05:13.395391Z", "rawValue": "6 pages, 2 figures. Accepted at the 2026 International Joint Conference on Neural Networks (IJCNN 2026), IEEE WCCI 2026; presented as an oral talk. Code and ART-SafeBench benchmark: https://github.com/FujitsuResearch/mirror", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2606.26793"}], "reviewStatus": "accepted", "venueName": "2026 International Joint Conference on Neural Networks (IJCNN 2026), IEEE WCCI 2026"}], "whyItMatters": "It provides a reusable cross-surface safety evaluation for agentic RAG systems rather than limiting testing to one attack channel."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
