{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.24589", "kind": "paper", "url": "https://arxiv.org/abs/2606.24589"}, {"id": "gh:khanak0509/adversabench", "kind": "repo", "url": "https://github.com/khanak0509/AdversaBench"}], "authors": ["Khanak Khandelwal"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.24589", "source_url": "https://arxiv.org/abs/2606.24589"}, "categories": ["Cybersecurity", "Language & Knowledge", "cs.AI"], "description": {"en": "AdversaBench is an automated LLM red-teaming evaluation that mutates seed prompts and confirms resulting failures with multiple judges and a tiebreaker."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.24589", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.24589", "title": "AdversaBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Indian Institute of Technology Jodhpur"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.24589", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.24589", "source_url": "https://openalex.org/W7165802700"}, "key": "claire-radar:2606.24589", "modality": null, "name": "AdversaBench", "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_adversabench_5199c227", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "1ae3e5d3df497578a31b191f8a6005063108fef2a047aba8967e25f35af90b7a", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T18:37:52.640057Z", "source_url": "https://arxiv.org/abs/2606.24589"}, "publisher": null, "released": "2026-06-23", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.24589", "source_url": "https://arxiv.org/abs/2606.24589"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-24589", "source": "claire_radar", "source_benchmark_id": "2606.24589", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Cybersecurity"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Automated red-teaming and cross-model transfer are active topics, while the reported experiment remains relatively small.", "score": 3}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T18:37:52.640057Z", "inputHash": "f756b1c90dd18a0c52040e7ed7a898c13e8c143b8e71ffda1ebb1d51c546fac5", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "AdversaBench: Automated LLM Red-Teaming with Multi-Judge Confirmation and Cross-Model Transferability", "canonicalNameSource": "paper_title", "decisionReason": "The source formally names the evaluation, describes its task and multi-judge failure-confirmation contract, and provides public code, data, and analysis scripts.", "evaluationEvidence": {"reuse": "Code, dataset, and analysis scripts are available at https://github.com/khanak0509/AdversaBench .", "scoring": "confirms failures through a three-judge panel with a meta-judge tiebreaker", "task": "reasoning, instruction-following, and tool use"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T18:37:52.640057Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "AdversaBench is an automated LLM red-teaming evaluation that mutates seed prompts and confirms resulting failures with multiple judges and a tiebreaker.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "public artifact URL"], "snippet": "We present AdversaBench, an end-to-end red-teaming pipeline that mutates seed prompts with five structured operators, queries a target model, and confirms failures through a three-judge panel with a meta-judge tiebreaker."}, "familyId": "bmf_df3e3661c1b6", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_adversabench_5199c227", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Cybersecurity"], "links": {"code": "https://github.com/khanak0509/AdversaBench", "data": null, "paper": "https://arxiv.org/abs/2606.24589", "pdf": "https://arxiv.org/pdf/2606.24589", "project": null, "report": "https://arxiv.org/abs/2606.24589"}, "metrics": [], "motivation": "Scaling adversarial evaluation of large language models requires both a method for generating hard inputs and a reliable way to confirm that resulting failures are real.", "name": "AdversaBench", "oneLine": "AdversaBench is an automated LLM red-teaming evaluation that mutates seed prompts and confirms resulting failures with multiple judges and a tiebreaker.", "paperTitle": "AdversaBench: Automated LLM Red-Teaming with Multi-Judge Confirmation and Cross-Model Transferability", "primaryDomain": "Cybersecurity", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.24589", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-06-23", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Khanak Khandelwal (Indian Institute of Technology Jodhpur)"], "categories": ["cs.AI", "cs.CL"], "id": "2606.24589", "title": "AdversaBench: Automated LLM Red-Teaming with Multi-Judge Confirmation and Cross-Model Transferability", "type": "arxiv", "url": "https://arxiv.org/abs/2606.24589"}, "sourceUpdatedAt": "2026-06-23T13:50:51Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "It supplies a repeatable way to compare adversarial failure discovery across reasoning, instruction-following, and tool-use tasks, including cross-model transfer."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
