{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.19396", "kind": "paper", "url": "https://arxiv.org/abs/2607.19396"}, {"kind": "repo", "url": "https://github.com/volkthienpreecha/crackedpdfs/releases/tag/v1.0.0-paper"}, {"id": "hf:volkthienpreecha/crackedpdfs", "kind": "dataset", "url": "https://huggingface.co/datasets/volkthienpreecha/crackedpdfs"}], "authors": ["Pukaphol Thienpreecha", "Karthik Subramanian"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.19396", "source_url": "https://arxiv.org/abs/2607.19396"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "Evaluates hidden prompt injection detection in PDFs through classification and paired ranking tasks, using 29,322 generated PDFs from 4,983 base documents. Includes frozen splits, features, and metrics for reproducibility."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.19396", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.19396", "title": "CrackedPDFs"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.19396", "modality": null, "name": "CrackedPDFs", "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": "unknown", "origin_record_id": "bm_crackedpdfs_1a52f043", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "b471702ef707f5ab406ff7f0dc11dc42dcfd9c360ce227d26aedd5ada47821e5", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.19396"}, "publisher": null, "released": "2026-07-03", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.19396", "source_url": "https://arxiv.org/abs/2607.19396"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-19396", "source": "claire_radar", "source_benchmark_id": "2607.19396", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/volkthienpreecha/crackedpdfs", "githubScope": "benchmark_repo", "githubStars": 2, "hfDailySubmittedAt": null, "hfDataset": "volkthienpreecha/crackedpdfs", "hfDatasetDownloads": 393, "hfDatasetLikes": 3, "hfPaperUpvotes": 0, "hfPaperUrl": "https://huggingface.co/papers/2607.19396", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}, "hfDatasetLikes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}}, "status": "fresh"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "b5aa65e925fc117358b2a0be920f582376b262ebdf182646d53138ebac700c78", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates hidden prompt injection detection in PDFs through classification and paired ranking tasks, using 29,322 generated PDFs from 4,983 base documents. Includes frozen splits, features, and metrics for reproducibility.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce CrackedPDFs, a controlled benchmark for hidden prompt injection in PDFs."}, "familyId": "bmf_e110c4f10728", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 2, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/volkthienpreecha/crackedpdfs/releases/tag/v1.0.0-paper", "url": "https://github.com/volkthienpreecha/crackedpdfs/releases/tag/v1.0.0-paper"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/volkthienpreecha/crackedpdfs", "url": "https://huggingface.co/papers/2607.19396"}, {"basis": "linked-from-project", "linkedUrl": "https://github.com/volkthienpreecha/crackedpdfs", "url": "https://doi.org/10.5281/zenodo.21735803"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/volkthienpreecha/crackedpdfs/releases/tag/v1.0.0-paper", "url": "https://arxiv.org/abs/2607.19396"}, {"basis": "linked-from-data", "linkedUrl": "https://github.com/volkthienpreecha/crackedpdfs/tree/main/tools/PDFautogenerator", "url": "https://huggingface.co/datasets/volkthienpreecha/crackedpdfs"}, {"basis": "linked-from-data", "linkedUrl": "https://github.com/volkthienpreecha/crackedpdfs.git", "url": "https://huggingface.co/datasets/volkthienpreecha/crackedpdfs"}, {"basis": "linked-from-data", "linkedUrl": "https://github.com/volkthienpreecha/crackedpdfs/tree/main/paper-v1", "url": "https://huggingface.co/datasets/volkthienpreecha/crackedpdfs"}, {"basis": "linked-from-data", "linkedUrl": "https://github.com/volkthienpreecha/crackedpdfs", "url": "https://huggingface.co/datasets/volkthienpreecha/crackedpdfs"}], "scope": "benchmark_repo", "url": "https://github.com/volkthienpreecha/crackedpdfs", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_crackedpdfs_1a52f043", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/volkthienpreecha/crackedpdfs/releases/tag/v1.0.0-paper", "data": "https://huggingface.co/datasets/volkthienpreecha/crackedpdfs", "hfPaper": "https://huggingface.co/papers/2607.19396", "paper": "https://arxiv.org/abs/2607.19396", "pdf": "https://arxiv.org/pdf/2607.19396", "project": "https://doi.org/10.5281/zenodo.21735803", "report": "https://arxiv.org/abs/2607.19396"}, "metrics": [], "motivation": "Document-based LLM systems often flatten a PDF before guardrails inspect it.", "name": "CrackedPDFs", "oneLine": "Evaluates hidden prompt injection detection in PDFs through classification and paired ranking tasks, using 29,322 generated PDFs from 4,983 base documents. Includes frozen splits, features, and metrics for reproducibility.", "paperTitle": "CrackedPDFs: A Controlled Benchmark for Hidden Prompt Injection in PDFs", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.19396", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-07-03", "source": {"authors": ["Pukaphol Thienpreecha (\"Volk\")", "Karthik Subramanian"], "categories": ["cs.AI"], "id": "2607.19396", "title": "CrackedPDFs: A Controlled Benchmark for Hidden Prompt Injection in PDFs", "type": "arxiv", "url": "https://arxiv.org/abs/2607.19396"}, "sourceUpdatedAt": "2026-08-02T14:07:20Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "Addresses the gap in evaluating defenses against prompt injections embedded in PDF structure, where flattening can hide malicious instructions. Provides a controlled paired benchmark with confounding controls to assess whether detectors generalize beyond superficial cues."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
