{"record": {"aliases": ["One Polluted Page Is Enough: Evaluating Web Content Pollution in Generative Recommenders"], "artifacts": [{"id": "arxiv:2606.13610", "kind": "paper", "url": "https://arxiv.org/abs/2606.13610"}, {"id": "gh:leoluolol/forge-benchmark", "kind": "repo", "url": "https://github.com/leoluolol/forge-benchmark"}], "authors": ["Minghao Luo", "Liang Chen"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.13610", "source_url": "https://arxiv.org/abs/2606.13610"}, "categories": ["Cybersecurity", "Language & Knowledge", "cs.CL"], "description": {"en": "FORGE evaluates whether search-augmented language models recommend fake products introduced through controlled modifications to retrieved web pages."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.13610", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.13610", "title": "FORGE"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.13610", "modality": null, "name": "FORGE", "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_one-polluted-page-is-enough-evaluating-web_01e33dfd", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "5111d0c996f086067a8aec4204c809aa17b2ce79e618df040c745ceb2d7bc23a", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T20:10:55.239724Z", "source_url": "https://arxiv.org/abs/2606.13610"}, "publisher": null, "released": "2026-06-11", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.13610", "source_url": "https://arxiv.org/abs/2606.13610"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-13610", "source": "claire_radar", "source_benchmark_id": "2606.13610", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["One Polluted Page Is Enough: Evaluating Web Content Pollution in Generative Recommenders"], "annotation": "Unknown", "applicationDomains": ["Cybersecurity"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses a timely safety risk in search-augmented recommendation systems and includes results across 12 models.", "score": 6}, "capabilities": [], "confidence": "Medium", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T20:10:55.239724Z", "inputHash": "e7685f591a33196bd4c0e74aec0e60b2b358c5c58d8530a5c73c7b578a62f467", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "we introduce FORGE (Fake Online Recommendations in Generative Environments)", "canonicalNameSource": "abstract", "decisionReason": "The source formally names the benchmark, defines controlled web-content pollution tasks, reports fooled-rate outcomes, and provides a public release repository.", "evaluationEvidence": {"reuse": "We release FORGE at https://github.com/leoluolol/forge-benchmark.", "scoring": "a single polluted page yields fooled rates of up to 27%, while the full top-3 replacement raises this to 73.8%", "task": "a benchmark for measuring fake-product promotion under controlled web-content pollution"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T20:10:55.239724Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "FORGE evaluates whether search-augmented language models recommend fake products introduced through controlled modifications to retrieved web pages.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "To answer this, we introduce FORGE (Fake Online Recommendations in Generative Environments), a benchmark for measuring fake-product promotion under controlled web-content pollution."}, "familyId": "bmf_ffc54a5fe1bb", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_one-polluted-page-is-enough-evaluating-web_01e33dfd", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Cybersecurity"], "links": {"code": "https://github.com/leoluolol/forge-benchmark", "data": null, "paper": "https://arxiv.org/abs/2606.13610", "pdf": "https://arxiv.org/pdf/2606.13610", "project": null, "report": "https://arxiv.org/abs/2606.13610"}, "metrics": [], "motivation": "Search-augmented LLMs increasingly mediate everyday consumer recommendations by retrieving live web content.", "name": "FORGE", "oneLine": "FORGE evaluates whether search-augmented language models recommend fake products introduced through controlled modifications to retrieved web pages.", "paperTitle": "One Polluted Page Is Enough: Evaluating Web Content Pollution in Generative Recommenders", "primaryDomain": "Cybersecurity", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.13610", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "introduces", "releasedAt": "2026-06-11", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Minghao Luo", "Liang Chen"], "categories": ["cs.CL", "cs.AI"], "id": "2606.13610", "title": "One Polluted Page Is Enough: Evaluating Web Content Pollution in Generative Recommenders", "type": "arxiv", "url": "https://arxiv.org/abs/2606.13610"}, "sourceUpdatedAt": "2026-06-11T17:24:14Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "It provides a repeatable way to compare the susceptibility of generative recommendation systems to polluted web content."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
