{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.23168", "kind": "paper", "url": "https://arxiv.org/abs/2605.23168"}], "authors": ["Luze Sun", "Anshuman Suri", "Harsh Chaudhari", "Cristina Nita-Rotaru", "Alina Oprea"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.23168", "source_url": "https://arxiv.org/abs/2605.23168"}, "categories": ["Cybersecurity", "Logistics & Operations", "Safety & Trustworthiness", "cs.CR"], "description": {"en": "PoisonForge benchmarks task-level targeted poisoning of instruction-tuned LLMs, parameterizing bias type, poisoning mode, appearance count, and target output length. It evaluates 12 open-weight models across five families with primarily 1% poison budget, using attack success rate as the main metric."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.23168", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.23168", "title": "PoisonForge"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Northeastern University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2605.23168", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2605.23168", "source_url": "https://openalex.org/W7162326943"}, "key": "claire-radar:2605.23168", "modality": null, "name": "PoisonForge", "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_poisonforge_b7c26795", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "3437d26da3438c61f349fb7fa56e832bc31105d27494a04ff71d0d89801d25ca", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2605.23168"}, "publisher": null, "released": "2026-05-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.23168", "source_url": "https://arxiv.org/abs/2605.23168"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-23168", "source": "claire_radar", "source_benchmark_id": "2605.23168", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Cybersecurity", "Logistics & Operations"], "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/2605.23168", "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": "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": "0b816d0d5409e4d3ebe05d45294a5e88bfd8eff3e42dcbb48db39ca9b9738f72", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "PoisonForge benchmarks task-level targeted poisoning of instruction-tuned LLMs, parameterizing bias type, poisoning mode, appearance count, and target output length. It evaluates 12 open-weight models across five families with primarily 1% poison budget, using attack success rate as the main metric.", "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"], "snippet": "We introduce PoisonForge, a benchmark that parameterizes this threat along four dimensions (bias type, poisoning mode, appearance count, and target output length) and evaluates 12 open-weight models (from 2B to 32B parameters) across five families under a primarily 1% poison budget."}, "familyId": "bmf_d4f13d62e72d", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_poisonforge_b7c26795", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Cybersecurity", "Logistics"], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2605.23168", "paper": "https://arxiv.org/abs/2605.23168", "pdf": "https://arxiv.org/pdf/2605.23168", "project": null, "report": "https://arxiv.org/abs/2605.23168"}, "metrics": [], "motivation": "When practitioners fine-tune LLMs on unvetted datasets, an adversary can exploit the data supply chain through task-level poisoning: inserting a small number of crafted instruction-response pairs that cause the model to embed attacker-specified entities, such as a country, in outputs for a targeted task family while behaving normally elsewhere.", "name": "PoisonForge", "oneLine": "PoisonForge benchmarks task-level targeted poisoning of instruction-tuned LLMs, parameterizing bias type, poisoning mode, appearance count, and target output length. It evaluates 12 open-weight models across five families with primarily 1% poison budget, using attack success rate as the main metric.", "paperTitle": "PoisonForge: Task-Level Targeted Poisoning Benchmark for Instruction-Tuned LLMs", "primaryDomain": "Cybersecurity", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.23168", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-22", "source": {"authors": ["Luze Sun", "Anshuman Suri", "Harsh Chaudhari", "Cristina Nita-Rotaru", "Alina Oprea"], "categories": ["cs.CR", "cs.AI", "cs.LG"], "id": "2605.23168", "title": "PoisonForge: Task-Level Targeted Poisoning Benchmark for Instruction-Tuned LLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2605.23168"}, "sourceUpdatedAt": "2026-05-22T02:41:13Z", "topics": ["cs.CR"], "venueAttempts": [], "whyItMatters": "Data supply chain poisoning poses a real threat when fine-tuning on unvetted data. PoisonForge quantifies vulnerability across models and configurations, highlighting that design choices rather than scale drive risk, aiding in risk assessment and mitigation."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
