{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.06477", "kind": "paper", "url": "https://arxiv.org/abs/2608.06477"}, {"id": "gh:borealisai/stepjack", "kind": "repo", "url": "https://github.com/BorealisAI/StepJack"}], "authors": ["Zhuoxin Zhan", "Akbar Rafiey", "Avery Ma", "Leila Pishdad", "Layla El Asri"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.06477", "source_url": "https://arxiv.org/abs/2608.06477"}, "categories": ["Agents", "General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "StepJack evaluates computer-use agents against multi-step indirect prompt injection attacks, where adversarial goals are decomposed into innocuous sub-steps across a chain of pages. It comprises 480 test examples across platforms and instruction types, with attack success rate as the primary metric."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.06477", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.06477", "title": "StepJack"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Borealis (Austria)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.06477", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.06477", "source_url": "https://openalex.org/W7202139345"}, "key": "claire-radar:2608.06477", "modality": null, "name": "StepJack", "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_stepjack_2287a7d0", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "7224f1ae2ed75c69517fb0a190c8756b9f68c504ad1632365512c90f4f247209", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.06477"}, "publisher": null, "released": "2026-08-06", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.06477", "source_url": "https://arxiv.org/abs/2608.06477"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-06477", "source": "claire_radar", "source_benchmark_id": "2608.06477", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/BorealisAI/StepJack", "githubScope": "benchmark_repo", "githubStars": 5, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.06477", "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"}, "capabilities": [], "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": "125383de0677db78ed7bbb12a15bc852e83fc2c788fd04f0283a6d6b919c4feb", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "StepJack evaluates computer-use agents against multi-step indirect prompt injection attacks, where adversarial goals are decomposed into innocuous sub-steps across a chain of pages. It comprises 480 test examples across platforms and instruction types, with attack success rate as the primary metric.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "With this pipeline, we build StepJack, a CUA safety benchmark with 480 test examples."}, "familyId": "bmf_43b09f4a2eb3", "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/BorealisAI/StepJack", "url": "https://github.com/BorealisAI/StepJack"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/BorealisAI/StepJack", "url": "https://huggingface.co/papers/2608.06477"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/BorealisAI/StepJack.", "url": "https://arxiv.org/abs/2608.06477"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/BorealisAI/StepJack", "url": "https://arxiv.org/abs/2608.06477"}], "scope": "benchmark_repo", "url": "https://github.com/BorealisAI/StepJack", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_stepjack_2287a7d0", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/BorealisAI/StepJack", "data": null, "hfPaper": "https://huggingface.co/papers/2608.06477", "paper": "https://arxiv.org/abs/2608.06477", "pdf": "https://arxiv.org/pdf/2608.06477", "project": null, "report": "https://arxiv.org/abs/2608.06477"}, "metrics": [], "motivation": "Computer-use agents (CUAs) face a growing threat from indirect prompt injection, where adversarial instructions are planted in the environment such as web pages.", "name": "StepJack", "oneLine": "StepJack evaluates computer-use agents against multi-step indirect prompt injection attacks, where adversarial goals are decomposed into innocuous sub-steps across a chain of pages. It comprises 480 test examples across platforms and instruction types, with attack success rate as the primary metric.", "paperTitle": "StepJack: Benchmarking Computer-Use Agent Safety Against Multi-Step Indirect Prompt Injection", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.06477", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "BorealisAI", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/BorealisAI/StepJack"}], "ranking": {"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": 0.2391676802567412, "value": 5}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 179, "score": 36}, "method": "attention-ranking-v13", "rank": 179, "score": 36}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-06", "source": {"authors": ["Zhuoxin Zhan", "Akbar Rafiey", "Avery Ma", "Leila Pishdad", "Layla El Asri"], "categories": ["cs.CR", "cs.AI", "cs.CL"], "id": "2608.06477", "title": "StepJack: Benchmarking Computer-Use Agent Safety Against Multi-Step Indirect Prompt Injection", "type": "arxiv", "url": "https://arxiv.org/abs/2608.06477"}, "sourceUpdatedAt": "2026-08-06T18:14:30Z", "topics": ["Agents", "Safety"], "venueAttempts": [], "whyItMatters": "The benchmark addresses the evaluation gap in agent safety against sophisticated, staged prompt injection attacks that previous single-step benchmarks fail to capture, offering a standardized way to assess and compare defenses for computer-use agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
