{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.06972", "kind": "paper", "url": "https://arxiv.org/abs/2609.06972"}, {"id": "gh:asif-0209/agentdrift", "kind": "repo", "url": "https://github.com/Asif-0209/AgentDrift"}], "authors": ["Asif Pinjari", "Mithun Paul Saint-Germain"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.06972", "source_url": "https://arxiv.org/abs/2609.06972"}, "categories": ["Agents", "General AI", "Language & Knowledge"], "description": {"en": "Step-labeled corpus of 12,536 synthetic LLM agent tool-call trajectories in five domains, labeled per step for benign, injection point, hijacked, or failed injection, supporting detection, localization, and attempt-vs-success evaluation."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.06972", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.06972", "title": "AgentDrift"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Northern Arizona University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.06972", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.06972", "source_url": "https://openalex.org/W7212034252"}, "key": "claire-radar:2609.06972", "modality": null, "name": "AgentDrift", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_agentdrift_563d3f60", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "72d08721ac3a80cbac25b73fb4204364ef2c36a3501470f3a484cbd36d8834d2", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-09T06:27:57.312318Z", "source_url": "https://arxiv.org/abs/2609.06972"}, "publisher": null, "released": "2026-09-07", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.06972", "source_url": "https://arxiv.org/abs/2609.06972"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-06972", "source": "claire_radar", "source_benchmark_id": "2609.06972", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/Asif-0209/AgentDrift", "githubScope": "benchmark_repo", "githubStars": 2, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2609.06972", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Security-focused step-level trajectory benchmark is topical and artifact-ready, likely to draw interest from agent safety researchers.", "score": 55}, "capabilities": [], "confidence": "High", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-09T06:27:57.312318Z", "inputHash": "a4c6ad561edd7599e81a89b4f6cc6ac1f99de9d02adf5dfbc68ebfa781060c79", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "AgentDrift: A Step-Labeled Benchmark of Injection-Hijacked LLM Agent Trajectories", "canonicalNameSource": "paper_title", "decisionReason": "Fully released dataset and evaluation protocols with clear metrics, made publicly available under an open license.", "evaluationEvidence": {"reuse": "The GitHub repository provides dataset files, splits, record schema, label grammar, and quick-start code, and the corpus is released under CC BY 4.0.", "scoring": "Tasks include detection (precision/recall/F1 on attacked class), localization (exact-match accuracy on injection index, IoU on hijacked span, macro step-level F1), and attempt vs. success (flag rate on failed attacks with recall per compliance pattern).", "task": "The corpus contains 12,536 synthetic tool-call trajectories over five agent domains in which every one of the 71,024 steps carries one of four labels: benign, injection point, hijacked, or failed injection."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-09T06:27:57.312318Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Step-labeled corpus of 12,536 synthetic LLM agent tool-call trajectories in five domains, labeled per step for benign, injection point, hijacked, or failed injection, supporting detection, localization, and attempt-vs-success evaluation.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We present AgentDrift, a benchmark of 12,536 synthetic tool-call trajectories over five agent domains in which every one of the 71,024 steps carries one of four labels: benign, injection point, hijacked, or failed injection."}, "familyId": "bmf_10f0ca733b7f", "firstSeenAt": "2026-09-09", "heat": null, "id": "bm_agentdrift_563d3f60", "indexedAt": "2026-09-09T06:18:29.794093Z", "industrySectors": [], "links": {"code": "https://github.com/Asif-0209/AgentDrift", "data": null, "paper": "https://arxiv.org/abs/2609.06972", "pdf": "https://arxiv.org/pdf/2609.06972", "project": null, "report": "https://arxiv.org/abs/2609.06972"}, "metrics": [], "motivation": "LLM agents complete tasks by issuing sequences of tool calls, and every observation they read is a channel through which an indirect prompt injection can enter.", "name": "AgentDrift", "oneLine": "Step-labeled corpus of 12,536 synthetic LLM agent tool-call trajectories in five domains, labeled per step for benign, injection point, hijacked, or failed injection, supporting detection, localization, and attempt-vs-success evaluation.", "paperTitle": "AgentDrift: A Step-Labeled Benchmark of Injection-Hijacked LLM Agent Trajectories", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.06972", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-09T06:18:29.794093Z"}, "publications": [], "publishers": [{"name": "Northern Arizona University", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/Asif-0209/AgentDrift"}], "ranking": {"30d": {"confidence": "Low", "coverage": 0.55, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.16198483906027114, "value": 2}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 80, "score": 31}, "method": "attention-ranking-v13", "rank": 80, "score": 31}, "90d": {"confidence": "Low", "coverage": 0.55, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.1466449917496476, "value": 2}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 253, "score": 31}, "method": "attention-ranking-v13", "rank": 253, "score": 31}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-09-07", "source": {"authors": ["Asif Pinjari", "Mithun Paul Saint-Germain"], "categories": ["cs.CR", "cs.AI", "cs.LG"], "id": "2609.06972", "title": "AgentDrift: A Step-Labeled Benchmark of Injection-Hijacked LLM Agent Trajectories", "type": "arxiv", "url": "https://arxiv.org/abs/2609.06972"}, "sourceUpdatedAt": "2026-09-07T03:10:23Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "Provides fine-grained trajectory labels for indirect prompt injection, enabling model development and evaluation beyond binary attack success."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
