{"record": {"aliases": ["Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security"], "artifacts": [{"id": "arxiv:2607.18063", "kind": "paper", "url": "https://arxiv.org/abs/2607.18063"}], "authors": ["Devina Jain", "David Hartmann", "Chuan Li"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.18063", "source_url": "https://arxiv.org/abs/2607.18063"}, "categories": ["Agents", "Agents & Tool Use", "Robotics & Embodied AI"], "description": {"en": "A 21-scenario benchmark evaluating memoryless LLM defenders against adaptive attackers that observe earlier responses and change tactics across multiple rounds."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.18063", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.18063", "title": "Adaptive Adversaries"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.18063", "modality": null, "name": "Adaptive Adversaries", "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_adaptive-adversaries-a-multi-turn-multi-ll_d2f62950", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "b9153eae0906c0450c916a5f20aabd85e891f84739eb45273dae59677f874fbe", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T19:14:05.917759Z", "source_url": "https://arxiv.org/abs/2607.18063"}, "publisher": null, "released": "2026-07-20", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.18063", "source_url": "https://arxiv.org/abs/2607.18063"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-18063", "source": "claire_radar", "source_benchmark_id": "2607.18063", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security"], "annotation": "Mixed", "applicationDomains": ["Robotics & Embodied AI"], "area": "Agents & Tool Use", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The multi-turn agent-security setting and comparisons among frontier models may attract near-term technical interest, although workshop positioning may limit broader reach.", "score": 6}, "capabilities": [], "confidence": "Medium", "construction": "Aggregate Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T19:13:08.979291Z", "inputHash": "6668b27a923bd97ebacf46ca1791e2e3623711327b096d00358086347f7c21a8", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security", "canonicalNameSource": "paper_title", "decisionReason": "Publish: the source formally names a benchmark, fixes scenarios and structured-output scoring, reports a comparable attack-success metric, and explicitly releases evaluation scenarios and runnable tooling.", "evaluationEvidence": {"reuse": "We release the benchmark -- 21 evaluation scenarios, 10 public development scenarios, the orchestrator, baseline harnesses, and a multi-attacker CLI", "scoring": "attack success rate (ASR)", "task": "adaptive multi-round attacks against memoryless LLM defenders"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T19:14:05.917759Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A 21-scenario benchmark evaluating memoryless LLM defenders against adaptive attackers that observe earlier responses and change tactics across multiple rounds.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We present a 21-scenario benchmark for \\emph{adaptive multi-round attacks against memoryless LLM defenders}: an autonomous LLM attacker observes prior defender responses and pivots across rounds, while each defender response is evaluated as a fresh interaction."}, "familyId": "bmf_8fdf81c076b0", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_adaptive-adversaries-a-multi-turn-multi-ll_d2f62950", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Robotics"], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2607.18063", "pdf": "https://arxiv.org/pdf/2607.18063", "project": null, "report": "https://arxiv.org/abs/2607.18063"}, "metrics": [], "motivation": "LLM-based agents process external content, exposing them to prompt injection and multi-turn manipulation.", "name": "Adaptive Adversaries", "oneLine": "A 21-scenario benchmark evaluating memoryless LLM defenders against adaptive attackers that observe earlier responses and change tactics across multiple rounds.", "paperTitle": "Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.18063", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-07-20", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Devina Jain", "David Hartmann", "Chuan Li"], "categories": ["cs.CR", "cs.AI", "cs.LG"], "id": "2607.18063", "title": "Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security", "type": "arxiv", "url": "https://arxiv.org/abs/2607.18063"}, "sourceUpdatedAt": "2026-07-20T15:30:38Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "It provides a repeatable protocol for measuring multi-round prompt-injection resilience and comparing defender performance using attack success rate."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
