{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.25707", "kind": "paper", "url": "https://arxiv.org/abs/2605.25707"}, {"id": "gh:tmlr-group/agenthijack", "kind": "repo", "url": "https://github.com/tmlr-group/AgentHijack"}, {"kind": "website", "url": "https://AgentHijack.github.io"}], "authors": ["Jingwei Sun", "Jianing Zhu", "Yuanyi Li", "Tongliang Liu", "Xia HU", "Bo Han"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.25707", "source_url": "https://arxiv.org/abs/2605.25707"}, "categories": ["Agents", "Agents & Tool Use", "General AI", "Robustness"], "description": {"en": "Evaluates the robustness of computer use agents under common environment corruptions such as pop-ups, resolution changes, and competing applications. The benchmark introduces 9 configurable corruptions and evaluates agent performance on desktop tasks using multimodal LLM-based agents, measuring task completion rates."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.25707", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.25707", "title": "AgentHijack"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2605.25707", "modality": null, "name": "AgentHijack", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_agenthijack_15cbe4d0", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "f0edfdf0cc026239d42d3425ede31286d4f111519b5fcf26ddc8623172fe928b", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2605.25707"}, "publisher": null, "released": "2026-05-25", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.25707", "source_url": "https://arxiv.org/abs/2605.25707"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-25707", "source": "claire_radar", "source_benchmark_id": "2605.25707", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Agents & Tool Use", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/tmlr-group/AgentHijack", "githubScope": "benchmark_repo", "githubStars": 7, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 3, "hfPaperUrl": "https://huggingface.co/papers/2605.25707", "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", "lastSuccessfulDate": "2026-09-30", "state": "stale"}}, "status": "fresh"}, "capabilities": ["Robustness"], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "8eec12423b8a25aee11bc508fa90f3d94f57674ede97c647b5518cf304495aa3", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates the robustness of computer use agents under common environment corruptions such as pop-ups, resolution changes, and competing applications. The benchmark introduces 9 configurable corruptions and evaluates agent performance on desktop tasks using multimodal LLM-based agents, measuring task completion rates.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce AgentHijack, a benchmark designed to evaluate the robustness of computer-use agents under common corruptions, where the uncertainties in dynamic environment disrupt the execution flow without direct adversarial intent."}, "familyId": "bmf_48bd013b2a40", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/tmlr-group/AgentHijack", "url": "https://github.com/tmlr-group/AgentHijack"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/tmlr-group/AgentHijack", "url": "https://huggingface.co/papers/2605.25707"}, {"basis": "huggingface-paper-association", "linkedUrl": "https://github.com/tmlr-group/AgentHijack", "url": "https://huggingface.co/api/papers/2605.25707"}, {"basis": "linked-from-project", "linkedUrl": "https://github.com/tmlr-group/AgentHijack", "url": "https://AgentHijack.github.io"}], "scope": "benchmark_repo", "url": "https://github.com/tmlr-group/AgentHijack", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_agenthijack_15cbe4d0", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/tmlr-group/AgentHijack", "data": null, "hfPaper": "https://huggingface.co/papers/2605.25707", "paper": "https://arxiv.org/abs/2605.25707", "pdf": "https://arxiv.org/pdf/2605.25707", "project": "https://AgentHijack.github.io", "report": "https://arxiv.org/abs/2605.25707"}, "metrics": [], "motivation": "Autonomous computer use agents that powered by multimodal large language models (MLLMs) are emerging as capable assistants for completing complex digital workflows.", "name": "AgentHijack", "oneLine": "Evaluates the robustness of computer use agents under common environment corruptions such as pop-ups, resolution changes, and competing applications. The benchmark introduces 9 configurable corruptions and evaluates agent performance on desktop tasks using multimodal LLM-based agents, measuring task completion rates.", "paperTitle": "AgentHijack: Benchmarking Computer Use Agent Robustness to Common Environment Corruptions", "primaryDomain": "General AI", "publication": {"evidence": "accepted by ICML 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2605.25707", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "ICML 2026", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "AgentHijack Team", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/tmlr-group/AgentHijack"}], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-25", "source": {"authors": ["Jingwei Sun and Jianing Zhu and Yuanyi Li and Tongliang Liu and Xia HU and Bo Han"], "categories": ["cs.AI"], "id": "2605.25707", "title": "AgentHijack: Benchmarking Computer Use Agent Robustness to Common Environment Corruptions", "type": "arxiv", "url": "https://arxiv.org/abs/2605.25707"}, "sourceUpdatedAt": "2026-05-25T11:09:22Z", "topics": ["Agents"], "venueAttempts": [{"decisionRaw": "accepted by ICML 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "accepted by ICML 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2605.25707"}], "reviewStatus": "accepted", "venueName": "ICML 2026"}], "whyItMatters": "Real-world execution environments are imperfect, and minor corruptions can cause significant performance degradation. This benchmark quantifies agent fragility and supports the development of more robust computer use agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
