{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.02684", "kind": "paper", "url": "https://arxiv.org/abs/2608.02684"}, {"id": "gh:pku-alignment/spike-bench", "kind": "repo", "url": "https://github.com/PKU-Alignment/SPIKE-Bench"}], "categories": ["Biology & Drug Discovery", "Safety", "Safety & Trustworthiness"], "description": {"en": "SPIKE-Bench evaluates LLM biosecurity risks using 631 curated toxin-design prompts and a three-stage funnel (compliance, plausibility, predicted toxicity) producing the Functional Harmfulness Rate."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.02684", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.02684", "title": "SPIKE-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.02684", "modality": null, "name": "SPIKE-Bench", "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_spike-bench_3b80754c", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "cfd52c3be169979f6c904106c00c35950e644841d919c9015cb2caee546d2bfb", "relation": "aggregates", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.02684"}, "publisher": null, "released": "2026-08-03", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.02684", "source_url": "https://arxiv.org/abs/2608.02684"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-02684", "source": "claire_radar", "source_benchmark_id": "2608.02684", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Mixed", "applicationDomains": ["Biology & Drug Discovery"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/PKU-Alignment/SPIKE-Bench", "githubScope": "benchmark_repo", "githubStars": 2, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.02684", "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": "Aggregate Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "f5a2312fc8ce8ce06a37a24ed20b5009d784e622db5fd68cc48c0356f8266367", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SPIKE-Bench evaluates LLM biosecurity risks using 631 curated toxin-design prompts and a three-stage funnel (compliance, plausibility, predicted toxicity) producing the Functional Harmfulness Rate.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "To address this evaluation blind spot, we introduce SPIKE-Bench, coupling 631 curated toxin-design prompts across seven functional categories with the SPIKE funnel, a three-stage protocol that filters output through compliance, biological plausibility, and predicted toxicity, producing stage-level diagnostics and an aggregate function-aware metric: the Functional Harmfulness Rate (FHR)."}, "familyId": "bmf_71b5d4d6ae47", "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/PKU-Alignment/SPIKE-Bench", "url": "https://github.com/PKU-Alignment/SPIKE-Bench"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/PKU-Alignment/SPIKE-Bench", "url": "https://huggingface.co/papers/2608.02684"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/PKU-Alignment/SPIKE-Bench", "url": "https://arxiv.org/abs/2608.02684"}], "scope": "benchmark_repo", "url": "https://github.com/PKU-Alignment/SPIKE-Bench", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_spike-bench_3b80754c", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Pharma & Biotech"], "links": {"code": "https://github.com/PKU-Alignment/SPIKE-Bench", "data": null, "hfPaper": "https://huggingface.co/papers/2608.02684", "paper": "https://arxiv.org/abs/2608.02684", "pdf": "https://arxiv.org/pdf/2608.02684", "project": null, "report": "https://arxiv.org/abs/2608.02684"}, "metrics": [], "motivation": "Large Language Models (LLMs) are accelerating biological research, yet this same capability poses a critical biosecurity threat: models that assist in protein engineering can equally be prompted to generate predicted toxin-like sequences, potentially lowering the barrier to biological misuse.", "name": "SPIKE-Bench", "oneLine": "SPIKE-Bench evaluates LLM biosecurity risks using 631 curated toxin-design prompts and a three-stage funnel (compliance, plausibility, predicted toxicity) producing the Functional Harmfulness Rate.", "paperTitle": "A Blind Spot in Alignment: Quantifying Biosecurity Risks in Large Language Models", "primaryDomain": "Biology & Drug Discovery", "publication": {"evidence": "Accepted to COLM 2026. 40 pages, 9 figures", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2608.02684", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "COLM 2026", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "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": 0.1466449917496476, "value": 2}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 268, "score": 31}, "method": "attention-ranking-v13", "rank": 268, "score": 31}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "aggregates", "releasedAt": "2026-08-03", "source": {"authors": ["Shu Quan", "Tianfang Hao", "Sitong Fang", "He Geng", "Jiayi Zhou", "Boyuan Chen", "Kaile Wang", "Donghai Hong", "Juntao Dai", "Yaodong Yang", "Jiaming Ji"], "categories": ["q-bio.QM", "cs.AI"], "id": "2608.02684", "title": "A Blind Spot in Alignment: Quantifying Biosecurity Risks in Large Language Models", "type": "arxiv", "url": "https://arxiv.org/abs/2608.02684"}, "sourceUpdatedAt": "2026-08-05T14:49:10Z", "topics": ["Safety"], "venueAttempts": [{"decisionRaw": "Accepted to COLM 2026. 40 pages, 9 figures", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted to COLM 2026. 40 pages, 9 figures", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2608.02684"}], "reviewStatus": "accepted", "venueName": "COLM 2026"}], "whyItMatters": "Provides a function-aware metric for biosecurity evaluation beyond refusal rates, enabling comparison of models on predicted functional harm."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
