{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.09404", "kind": "paper", "url": "https://arxiv.org/abs/2609.09404"}], "authors": ["Viet K. Nguyen", "Mohammad I. Husain"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.09404", "source_url": "https://arxiv.org/abs/2609.09404"}, "categories": ["General AI", "Multimodal"], "description": {"en": "A benchmark for evaluating multimodal prompt injection attacks on agentic AI frameworks."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.09404", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.09404", "title": "MMPIBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["California State Polytechnic University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.09404", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.09404", "source_url": "https://openalex.org/W7212157363"}, "key": "claire-radar:2609.09404", "modality": null, "name": "MMPIBench", "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-indexed", "display_eligible": "false", "origin_record_id": "bm_mmpibench_8b6b75b3", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "e7215c181ed54c89545100ca2187421041859f98485f371ca967148e31f83064", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-12T00:56:20.009268Z", "source_url": "https://arxiv.org/abs/2609.09404"}, "publisher": null, "released": "2026-09-08", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.09404", "source_url": "https://arxiv.org/abs/2609.09404"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-09404", "source": "claire_radar", "source_benchmark_id": "2609.09404", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Multimodal", "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/2609.09404", "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": "Prompt injection is a hot security topic, but lack of released artifacts may cap immediate engagement.", "score": 40}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-12T00:56:20.009268Z", "inputHash": "0236694b8e594cdb6362aa9b895cc659e6dbf6f3fed35dc939a24234a9d22aef", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We present MMPIBench, a reproducible benchmark that measures what happens next.", "canonicalNameSource": "abstract", "decisionReason": "No artifact evidence, code, or public reuse path is supplied, so the benchmark's content and scoring method cannot be confirmed.", "evaluationEvidence": {"reuse": "MMPIBench is a reproducible benchmark that can be used to evaluate multimodal prompt injection attacks across agentic AI frameworks.", "scoring": "Across 720 runs covering six frameworks, five foundation models, six carriers, and four attacker objectives, attacks complete in approximately 1% of runs but are attempted in 12.8%.", "task": "MMPIBench delivers a fixed set of attacks through six visual carriers (OCR text, overlays, EXIF metadata, QR codes, fake interfaces, and hybrids) and records how far each injected instruction travels through the agent, from perception through planning to the tool call."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-12T00:56:20.009268Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A benchmark for evaluating multimodal prompt injection attacks on agentic AI frameworks.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence"], "snippet": "We present MMPIBench, a reproducible benchmark that measures what happens next."}, "familyId": "bmf_58cd860387d8", "firstSeenAt": "2026-09-10", "heat": null, "id": "bm_mmpibench_8b6b75b3", "indexedAt": "2026-09-10T00:47:26.889471Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.09404", "pdf": "https://arxiv.org/pdf/2609.09404", "project": null, "report": "https://arxiv.org/abs/2609.09404"}, "metrics": [], "motivation": "Agentic AI frameworks let a language model plan, keep memory, and call tools that reach real files, mail, and services.", "name": "MMPIBench", "oneLine": "A benchmark for evaluating multimodal prompt injection attacks on agentic AI frameworks.", "paperTitle": "An Experimental Evaluation of Multimodal Prompt Injection Attacks on Agentic AI Frameworks", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.09404", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-10T00:47:26.889471Z"}, "publications": [], "ranking": {"30d": {"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": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "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}, "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": 0.85, "relation": "introduces", "releasedAt": "2026-09-08", "source": {"authors": ["Viet K. Nguyen", "Mohammad I. Husain"], "categories": ["cs.CR", "cs.AI"], "id": "2609.09404", "title": "An Experimental Evaluation of Multimodal Prompt Injection Attacks on Agentic AI Frameworks", "type": "arxiv", "url": "https://arxiv.org/abs/2609.09404"}, "sourceUpdatedAt": "2026-09-08T20:00:03Z", "topics": ["Multimodal"], "venueAttempts": [], "whyItMatters": "Security of agentic systems is a growing concern, and reproducible benchmarks can help compare defenses."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
