{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.02240", "kind": "paper", "url": "https://arxiv.org/abs/2606.02240"}], "authors": ["Hiskias Dingeto", "William Leeney"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.02240", "source_url": "https://arxiv.org/abs/2606.02240"}, "categories": ["General AI", "Language & Knowledge", "cs.CR"], "description": {"en": "AgentRedBench evaluates LLM agents against indirect prompt injection and underspecified-authorization attacks across 24 enterprise SaaS integrations. It defines 215 attack scenarios with immutable versioning, and tracks attack success rate (ASR) for models and defenses. Open-source codebase and schemas enable replay."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.02240", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.02240", "title": "AgentRedBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.02240", "modality": null, "name": "AgentRedBench", "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_agentredbench_2025428a", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "3d23bc0ba0bdcf07519f3575ff430d0531edc21934338e8775454097b1eec5f6", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.02240"}, "publisher": null, "released": "2026-06-01", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.02240", "source_url": "https://arxiv.org/abs/2606.02240"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-02240", "source": "claire_radar", "source_benchmark_id": "2606.02240", "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": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2606.02240", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "capabilities": [], "confidence": "High", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "589fc791f1a241604065026b182fcb2f17bb478b63f04e4d447c9ec10588c3de", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "AgentRedBench evaluates LLM agents against indirect prompt injection and underspecified-authorization attacks across 24 enterprise SaaS integrations. It defines 215 attack scenarios with immutable versioning, and tracks attack success rate (ASR) for models and defenses. Open-source codebase and schemas enable replay.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce AGENTREDBENCH, a dynamic LLM-driven redteaming benchmark of 215 subtle underspecified-authorization scenarios across 24 enterprise integrations and five attack types."}, "familyId": "bmf_1ff4fb25b17b", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_agentredbench_2025428a", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.02240", "paper": "https://arxiv.org/abs/2606.02240", "pdf": "https://arxiv.org/pdf/2606.02240", "project": null, "report": "https://arxiv.org/abs/2606.02240"}, "metrics": [], "motivation": "Indirect prompt injection in tool-use agents is a concrete production threat: LLM agents read from integrations (third-party services such as Gmail, Salesforce, or Jira accessed through tool calls) whose response content the user neither writes nor controls.", "name": "AgentRedBench", "oneLine": "AgentRedBench evaluates LLM agents against indirect prompt injection and underspecified-authorization attacks across 24 enterprise SaaS integrations. It defines 215 attack scenarios with immutable versioning, and tracks attack success rate (ASR) for models and defenses. Open-source codebase and schemas enable replay.", "paperTitle": "AgentRedBench: Dynamic Redteaming and Integration-Aware Defense for LLM Agents over SaaS Integrations", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.02240", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "arXiv", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://arxiv.org/abs/2606.02240"}], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-01", "source": {"authors": ["Hiskias Dingeto", "William Leeney"], "categories": ["cs.CR", "cs.AI", "cs.CL", "cs.ET"], "id": "2606.02240", "title": "AgentRedBench: Dynamic Redteaming and Integration-Aware Defense for LLM Agents over SaaS Integrations", "type": "arxiv", "url": "https://arxiv.org/abs/2606.02240"}, "sourceUpdatedAt": "2026-07-17T12:28:54Z", "topics": ["cs.CR"], "venueAttempts": [], "whyItMatters": "Prior agent-security benchmarks cover limited integrations with static payloads, understating real-world exploitability. AgentRedBench provides a dynamic, maintainer-mediated evaluation to compare model and guard-rail effectiveness against evolving injection threats, supporting procurement and hardening decisions."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
