{"record": {"aliases": ["Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents"], "artifacts": [{"id": "arxiv:2608.12851", "kind": "paper", "url": "https://arxiv.org/abs/2608.12851"}, {"id": "gh:henrymao2004/misevolve", "kind": "repo", "url": "https://github.com/henrymao2004/misevolve"}], "authors": ["Xutao Mao", "Liangjie Zhao", "Xiang Zheng", "Cong Wang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.12851", "source_url": "https://arxiv.org/abs/2608.12851"}, "categories": ["Agents", "General AI", "Language & Knowledge"], "description": {"en": "Lifecycle-aware benchmark for persistent safety failures in self-improving LLM agents, consisting of 25 episodes with malicious demonstrations, benign twins, and fresh-session probes, scored with nine metrics including carryover ASR and unsafe retrieval."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.12851", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.12851", "title": "SkillMisevo-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["City University of Hong Kong", "Adelaide University", "The University of Adelaide"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.12851", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.12851", "source_url": "https://openalex.org/W7203492321"}, "key": "claire-radar:2608.12851", "modality": null, "name": "SkillMisevo-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "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_practice-makes-unsafe-skill-misevolution-i_3d2c225f", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "280f75b8ddfe0fef94702be1754f16ffecfef92aaff91395a220296e86962ca5", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-08-26T06:10:57.123445Z", "source_url": "https://arxiv.org/abs/2608.12851"}, "publisher": null, "released": "2026-08-13", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.12851", "source_url": "https://arxiv.org/abs/2608.12851"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-12851", "source": "claire_radar", "source_benchmark_id": "2608.12851", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/henrymao2004/misevolve", "githubScope": "benchmark_repo", "githubStars": 7, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.12851", "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": "The safety-focused angle and provided code repository should drive early interest among agent safety researchers.", "score": 58}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-26T06:10:57.123445Z", "inputHash": "785db45ce0c88fa23dc05edc0e20cfda2a742d55ee5cf0d7b8844b696afecb2b", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "SkillMisevo-Bench, a frozen design from malicious exposure to carryover tasks, with concept-aligned benign tasks and nine lifecycle metrics", "canonicalNameSource": "abstract", "decisionReason": "The frozen benchmark design with clear episodes and metrics provides a stable scoring contract, and the repository includes code and evaluation procedures for public reuse.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-26T06:10:57.123445Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Lifecycle-aware benchmark for persistent safety failures in self-improving LLM agents, consisting of 25 episodes with malicious demonstrations, benign twins, and fresh-session probes, scored with nine metrics including carryover ASR and unsafe retrieval.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "To expose this lifecycle, we introduce SkillMisevo-Gym, a lifecycle-aware harness that versions skill state across agent frameworks, and SkillMisevo-Bench, a frozen design from malicious exposure to carryover tasks, with concept-aligned benign tasks and nine lifecycle metrics."}, "familyId": "bmf_be0110c7fe59", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_practice-makes-unsafe-skill-misevolution-i_3d2c225f", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/henrymao2004/misevolve", "data": null, "paper": "https://arxiv.org/abs/2608.12851", "pdf": "https://arxiv.org/pdf/2608.12851", "project": null, "report": "https://arxiv.org/abs/2608.12851"}, "metrics": [], "motivation": "Self-improving LLM agents convert successful trajectories into persistent cross-task state.", "name": "SkillMisevo-Bench", "oneLine": "Lifecycle-aware benchmark for persistent safety failures in self-improving LLM agents, consisting of 25 episodes with malicious demonstrations, benign twins, and fresh-session probes, scored with nine metrics including carryover ASR and unsafe retrieval.", "paperTitle": "Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.12851", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "Henry Mao and collaborators", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/henrymao2004/misevolve"}], "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": 0.2775680655212807, "value": 7}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 159, "score": 38}, "method": "attention-ranking-v13", "rank": 159, "score": 38}}, "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-08-13", "source": {"authors": ["Xutao Mao and Liangjie Zhao and Xiang Zheng and Cong Wang"], "categories": ["cs.AI"], "id": "2608.12851", "title": "Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2608.12851"}, "sourceUpdatedAt": "2026-08-13T05:47:43Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "Exposes how unsafe experiences can become reusable policies that cause later harm, enabling measurement of risk across skill authoring, retrieval, and reuse."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
