{"record": {"aliases": ["From Prompt Injection to Persistent Control: Defending Agentic Harness Against Trojan Backdoors"], "artifacts": [{"id": "arxiv:2605.31042", "kind": "paper", "url": "https://arxiv.org/abs/2605.31042"}, {"id": "gh:ruc-nlpir/clawtrojan", "kind": "repo", "url": "https://github.com/RUC-NLPIR/ClawTrojan"}], "authors": ["Jiejun Tan", "Zhicheng Dou", "Xinyu Yang", "Yuyang Hu", "Yiruo Cheng", "Xiaoxi Li", "Ji-Rong Wen"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.31042", "source_url": "https://arxiv.org/abs/2605.31042"}, "categories": ["General AI", "Language & Knowledge", "cs.CR"], "description": {"en": "A reusable benchmark for measuring multi-step trojan attacks that establish persistent control through local agent workspace state."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.31042", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.31042", "title": "ClawTrojan"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Renmin University of China"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2605.31042", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2605.31042", "source_url": "https://openalex.org/W7162993931"}, "key": "claire-radar:2605.31042", "modality": null, "name": "ClawTrojan", "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_from-prompt-injection-to-persistent-contro_f6658463", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "e5cb07fc86392747b0e738b93ef904b54aff20f865ed72185567c7847ee9e2f4", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T19:47:04.170876Z", "source_url": "https://arxiv.org/abs/2605.31042"}, "publisher": null, "released": "2026-05-29", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.31042", "source_url": "https://arxiv.org/abs/2605.31042"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-31042", "source": "claire_radar", "source_benchmark_id": "2605.31042", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["From Prompt Injection to Persistent Control: Defending Agentic Harness Against Trojan Backdoors"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Persistent attacks against tool-using agents are a timely security concern, and the available code and data make immediate examination feasible.", "score": 4}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T19:47:04.170876Z", "inputHash": "02535fb28256bec8328113f4998656441513e7531eb3b8f47385199c497cc973", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "we introduce ClawTrojan, a benchmark", "canonicalNameSource": "abstract", "decisionReason": "The source formally names the benchmark, defines its attack-detection task, reports attack success rate as a concrete outcome metric, and explicitly makes code and data available through a public link.", "evaluationEvidence": {"reuse": "Code and data are available at https://github.com/RUC-NLPIR/ClawTrojan", "scoring": "ClawTrojan reaches a 95.5% attack success rate (ASR)", "task": "designed to identify multi-step trojan attacks in local agentic harnesses"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T19:47:04.170876Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A reusable benchmark for measuring multi-step trojan attacks that establish persistent control through local agent workspace state.", "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 reveal this threat, we introduce ClawTrojan, a benchmark designed to identify multi-step trojan attacks in local agentic harnesses."}, "familyId": "bmf_4ceabbf82411", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_from-prompt-injection-to-persistent-contro_f6658463", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/RUC-NLPIR/ClawTrojan", "data": null, "paper": "https://arxiv.org/abs/2605.31042", "pdf": "https://arxiv.org/pdf/2605.31042", "project": null, "report": "https://arxiv.org/abs/2605.31042"}, "metrics": [], "motivation": "LLM agents are evolving from conversational chatbots to operational tools in real-world workspaces.", "name": "ClawTrojan", "oneLine": "A reusable benchmark for measuring multi-step trojan attacks that establish persistent control through local agent workspace state.", "paperTitle": "From Prompt Injection to Persistent Control: Defending Agentic Harness Against Trojan Backdoors", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.31042", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "RUC-NLPIR", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/RUC-NLPIR/ClawTrojan"}], "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "introduces", "releasedAt": "2026-05-29", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Jiejun Tan", "Zhicheng Dou", "Xinyu Yang", "Yuyang Hu", "Yiruo Cheng", "Xiaoxi Li", "Ji-Rong Wen"], "categories": ["cs.CR", "cs.AI", "cs.CL"], "id": "2605.31042", "title": "From Prompt Injection to Persistent Control: Defending Agentic Harness Against Trojan Backdoors", "type": "arxiv", "url": "https://arxiv.org/abs/2605.31042"}, "sourceUpdatedAt": "2026-05-29T09:19:07Z", "topics": ["cs.CR"], "venueAttempts": [], "whyItMatters": "It evaluates attacks whose individually benign-looking steps can evade defenses that inspect actions in isolation."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
