{"record": {"aliases": ["When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models"], "artifacts": [{"id": "arxiv:2605.27851", "kind": "paper", "url": "https://arxiv.org/abs/2605.27851"}], "authors": ["Dasol Choi", "Alex Kwon"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.27851", "source_url": "https://arxiv.org/abs/2605.27851"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "Context-flip evaluation tests whether aligned language models adjust their actions when a situational change reverses which action is safe."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.27851", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.27851", "title": "context-flip evaluation"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2605.27851", "modality": null, "name": "context-flip evaluation", "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_when-context-flips-safety-breaks-diagnosin_f094d571", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "16c275e3c4f5bae604b66fcdeeae38b34b1cbf556e1b2fe6105405f98caa66aa", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T18:01:34.017151Z", "source_url": "https://arxiv.org/abs/2605.27851"}, "publisher": null, "released": "2026-05-27", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.27851", "source_url": "https://arxiv.org/abs/2605.27851"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-27851", "source": "claire_radar", "source_benchmark_id": "2605.27851", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "The released protocol addresses a concrete limitation of conventional language-model safety scores.", "score": 3}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T18:01:34.017151Z", "inputHash": "ef9696c642b3d8f0445449f4d690b11a1fd3e51decdb111d383f33b57e083e24", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "we introduce context-flip evaluation", "canonicalNameSource": "abstract", "decisionReason": "The abstract formally names a repeatable evaluation, describes paired evaluation cases and a comparable score, and states that the protocol and perturbed benchmarks are released.", "evaluationEvidence": {"reuse": "We release our protocol, perturbed benchmarks, and deployment probe.", "scoring": "all 12 models exhibit a safety-commonsense gap (mean +17.4 pp)", "task": "testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T18:01:34.017151Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Context-flip evaluation tests whether aligned language models adjust their actions when a situational change reverses which action is safe.", "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"], "snippet": "To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm."}, "familyId": "bmf_4a430e20484e", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_when-context-flips-safety-breaks-diagnosin_f094d571", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2605.27851", "pdf": "https://arxiv.org/pdf/2605.27851", "project": null, "report": "https://arxiv.org/abs/2605.27851"}, "metrics": [], "motivation": "Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe.", "name": "context-flip evaluation", "oneLine": "Context-flip evaluation tests whether aligned language models adjust their actions when a situational change reverses which action is safe.", "paperTitle": "When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.27851", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-05-27", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Dasol Choi", "Alex Kwon"], "categories": ["cs.AI"], "id": "2605.27851", "title": "When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models", "type": "arxiv", "url": "https://arxiv.org/abs/2605.27851"}, "sourceUpdatedAt": "2026-05-27T02:09:30Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "It measures safety brittleness that can remain hidden behind high baseline benchmark accuracy."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
