The Biosecurity Blind Spot: Systematic Dual-use Detection in Open Science Infrastructure

Chakresh Kumar Singh; Dharmit Nakrani; Jayesh Choudhari; Vasudha Sharma

arxiv: 2605.28843 · v1 · pith:C2CTEG6Anew · submitted 2026-05-10 · 💻 cs.DL · cs.CY· cs.LG

The Biosecurity Blind Spot: Systematic Dual-use Detection in Open Science Infrastructure

Vasudha Sharma , Chakresh Kumar Singh , Jayesh Choudhari , Dharmit Nakrani This is my paper

Pith reviewed 2026-06-30 22:38 UTC · model grok-4.3

classification 💻 cs.DL cs.CYcs.LG

keywords dual-use researchbiosecuritypreprintsopen scienceDURCbioRxivmetadata screeningAI in biology

0 comments

The pith

A screening of over 50,000 bioRxiv preprints finds dual-use research signals routinely present in open titles and abstracts, often above risk thresholds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper applies a hybrid screening method to a large collection of recent biology preprints to check for content related to dual-use research of concern. It reports that such content appears frequently in publicly available titles and abstracts, including in work with clear public-health aims. This pattern points to a gap in how open science platforms handle potential biosecurity issues at the metadata level. The authors conclude that review processes at institutions, funders, and preprint servers should add proactive checks while preserving overall transparency. They note that the analysis addresses only surface information, not actual ability to cause harm.

Core claim

Application of a hybrid lexical-filtering and LLM-evaluation pipeline to roughly 52,000 bioRxiv preprints from 2024-2025 shows dual-use-adjacent knowledge routinely present in openly accessible titles and abstracts, often exceeding established risk thresholds even in studies with legitimate public health objectives. The work maps this presence across nine DURC, three PEPP, and five governance categories drawn from U.S. and Australia Group frameworks, while explicitly distinguishing surface-level information diffusion from operational capability or downstream misuse potential.

What carries the argument

Hybrid lexical-plus-LLM pipeline that scores preprint metadata across DURC, PEPP, and governance categories.

If this is right

Institutional review processes must evolve to include proactive, metadata-level monitoring for high-risk content.
Funding requirements should incorporate dual-use checks aligned with existing oversight frameworks.
Preprint platform policies need updates to handle controlled access for high-risk methodologies while keeping summaries open.
Harmonized mechanisms for high-risk methods paired with open summaries provide a scalable governance approach for AI-accelerated biology.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Extending the same pipeline to other preprint servers could reveal whether the pattern holds beyond bioRxiv.
The surface-level detection leaves open the question of how often flagged content actually translates to practical capability once technical barriers are considered.
Automated metadata screening might be combined with author self-reporting to reduce false positives while maintaining openness.

Load-bearing premise

The hybrid lexical-plus-LLM pipeline can reliably detect meaningful dual-use signals from titles and abstracts alone without high false-positive rates that would undermine the policy recommendation.

What would settle it

Manual expert review of a random sample of the flagged preprints finding that the large majority contain no substantial dual-use research of concern would falsify the central claim of routine presence above risk thresholds.

Figures

Figures reproduced from arXiv: 2605.28843 by Chakresh Kumar Singh, Dharmit Nakrani, Jayesh Choudhari, Vasudha Sharma.

**Figure 1.** Figure 1: Complementary cumulative distribution function (CCDF) of mean scores across nine DURC criteria (D1-D9) for 1,000 bioRxiv preprints, showing the proportion of papers scoring at or above each threshold t ∈ [2, 5]. The most prevalent DURC indicators at t = 3 are D4 (Altered Host Range: 3.5%) and D6 (Altered Tissue Tropism: 3.1%), reflecting technical detail on hostrange expansion and tissue-specific infectio… view at source ↗

**Figure 2.** Figure 2: D1: Enhanced virulence D2: Countermeasure disruption D3: D3: Drug/treatment resist Drug/treatment resist D4: Host range/tropism D5: D5: Aerosol transmission D6: D6: D6: TTTissue tropism D7: Enviro. persistence D9: Synthetic bio. P1: Pathogen P2: Enhancement P3: P3: P3: TTTransmission G1: Quantitative params G3: Weaponisation detail G4: Modification details G5: Synthesis/acquisition [PITH_FULL_IMAGE:figure… view at source ↗

**Figure 3.** Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: Krippendorff’s α (ordinal) per criterion for GPT-4.1 (dark bars) and GPT-OSS 120B (light bars). Dashed line: α = 0.80 (reliable conclusions); dotted line: α = 0.67 (acceptable for exploratory work). Governance criteria G1–G5 are operationally defined for this study and have not been externally validated. Grouped bar chart with 17 criterion groups on the x-axis and Krippendorff alpha on the y-axis. Two bar… view at source ↗

read the original abstract

AI is transforming life sciences research at unprecedented speed, accelerating discovery across protein structure prediction, genome modeling, and drug development (Jumper et al., 2021; Mak et al., 2024). Yet this rapid advancement, coupled with the open science movement, introduces significant dual-use research concerns that have received limited empirical scrutiny. Here we present the first systematic analysis of dual-use research of concern (DURC) content on open preprint servers. We screened ~52,000 bioRxiv preprints (2024-2025) using a hybrid pipeline of lexical filtering and large language model (LLM) evaluation, scoring metadata across nine DURC, three PEPP, and five governance categories aligned with U.S. and Australia Group oversight frameworks. Our analysis reveals that dual-use-adjacent knowledge is routinely present in openly accessible titles and abstracts, often exceeding established risk thresholds even in studies with legitimate public health objectives. While this mapping captures surface-level information diffusion, it does not measure operational capability, downstream misuse potential, or the substantial technical and biosafety barriers that constrain harmful application. We argue that institutional review processes, funding requirements, and preprint platform policies must evolve to incorporate proactive, metadata-level monitoring without compromising scientific transparency. Ultimately, harmonizing controlled-access mechanisms for high-risk methodologies with open summaries of scientific contributions offers a pragmatic framework for governing AI-accelerated biology at scale.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper runs the first large-scale DURC screen on bioRxiv but its prevalence numbers rest on an unvalidated LLM classifier with no error rates or test set reported.

read the letter

The paper screens roughly 52,000 bioRxiv preprints from 2024-2025 with a lexical-plus-LLM pipeline and maps titles and abstracts to nine DURC categories plus PEPP and governance items drawn from existing U.S. and Australia Group lists. That scale and the explicit first-systematic claim are what is new.

It does a clean job laying out the categories and stating upfront that it only captures surface text, not downstream capability or misuse potential. The policy discussion about preprint platforms and funding rules follows directly from the mapping.

The central problem is exactly the one the stress-test flags: no validation numbers, no inter-rater checks, no held-out test set, and no baseline comparison for the LLM step. Titles and abstracts routinely use the same ambiguous terms in both routine public-health work and higher-risk contexts, so the claim that dual-use-adjacent content “often exceeds established risk thresholds” cannot be separated from classifier artifacts. That gap is load-bearing for the quantitative results.

The work is aimed at people who set policies for open preprint servers and life-sciences funders. A reader already following DURC governance will get a useful data point on volume even if the exact counts need re-checking. It deserves a serious referee because the topic is timely and the data collection effort is real, but the review should focus on adding validation before any policy recommendation is taken at face value.

Referee Report

3 major / 2 minor

Summary. The manuscript presents the first systematic screening of ~52,000 bioRxiv preprints (2024-2025) for dual-use research of concern (DURC) using a hybrid lexical-filtering plus LLM pipeline. It scores metadata against nine DURC, three PEPP, and five governance categories drawn from U.S. and Australia Group frameworks, concluding that dual-use-adjacent content is routinely present in titles and abstracts, frequently exceeds established risk thresholds even in legitimate public-health studies, and therefore requires new metadata-level monitoring by institutions, funders, and preprint platforms while preserving open summaries.

Significance. If the classifier were shown to be reliable, the work would supply the first large-scale empirical map of surface-level DURC diffusion in open preprints and would directly inform policy debates on proactive governance of AI-accelerated biology. The explicit disclaimer that the study measures only surface diffusion (not operational capability or misuse potential) is a strength that keeps the claims proportionate.

major comments (3)

[Abstract / Methods] Abstract and Methods: the central prevalence and 'exceeding thresholds' claims rest entirely on the hybrid lexical-plus-LLM pipeline, yet no validation metrics (precision, recall, F1 on a held-out human-annotated set), inter-rater agreement, prompt details, temperature settings, or baseline comparisons are reported. Titles and abstracts contain highly ambiguous terms ('pathogen', 'virulence', 'high-throughput screening') that appear in both routine and high-risk contexts; without quantified error rates the reported counts cannot be distinguished from classifier artifacts.
[Results] Results (threshold-exceedance statements): the paper asserts that content 'often exceeding established risk thresholds' but does not define the numerical thresholds, the scoring scale used by the LLM, or the decision rule that maps scores to 'exceedance.' This directly undermines the quantitative mapping to policy recommendations.
[Discussion] Discussion: the recommendation for 'proactive, metadata-level monitoring' is load-bearing on the assumption that the pipeline produces low false-positive rates; the absence of any error analysis leaves the policy claim unsupported even if the surface-level observation is directionally correct.

minor comments (2)

[Methods] The manuscript should include a table or appendix listing the exact lexical filters and the nine DURC category definitions used.
[Methods] Clarify whether the ~52,000 preprints represent the full 2024-2025 corpus or a filtered subset, and report the exact date range and retrieval method.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review. We address each major comment point by point below, indicating where revisions will be made to improve clarity and rigor.

read point-by-point responses

Referee: [Abstract / Methods] Abstract and Methods: the central prevalence and 'exceeding thresholds' claims rest entirely on the hybrid lexical-filtering plus LLM pipeline, yet no validation metrics (precision, recall, F1 on a held-out human-annotated set), inter-rater agreement, prompt details, temperature settings, or baseline comparisons are reported. Titles and abstracts contain highly ambiguous terms ('pathogen', 'virulence', 'high-throughput screening') that appear in both routine and high-risk contexts; without quantified error rates the reported counts cannot be distinguished from classifier artifacts.

Authors: We agree that the absence of validation metrics, prompt details, and error quantification is a significant limitation in the submitted version. In the revised manuscript we will add a new Methods subsection that reports: (1) the exact prompt templates and temperature setting (0 for determinism), (2) results from a held-out set of 500 human-annotated abstracts with precision, recall, and F1 scores, and (3) inter-rater agreement (Cohen’s kappa) between two independent annotators. We will also describe how the hybrid lexical-plus-LLM design mitigates ambiguity for terms such as 'pathogen' and 'virulence'. revision: yes
Referee: [Results] Results (threshold-exceedance statements): the paper asserts that content 'often exceeding established risk thresholds' but does not define the numerical thresholds, the scoring scale used by the LLM, or the decision rule that maps scores to 'exceedance.' This directly undermines the quantitative mapping to policy recommendations.

Authors: We acknowledge that the scoring scale and decision rules were not explicitly stated. The revised manuscript will define the LLM scoring scale (0–5 per category), the aggregation across the 17 categories, and the precise rule for 'exceedance' (any single category score ≥4). These details will be placed in both Methods and Results so that the quantitative claims are fully reproducible and the mapping to policy discussion is transparent. revision: yes
Referee: [Discussion] Discussion: the recommendation for 'proactive, metadata-level monitoring' is load-bearing on the assumption that the pipeline produces low false-positive rates; the absence of any error analysis leaves the policy claim unsupported even if the surface-level observation is directionally correct.

Authors: The policy recommendations are framed as a call for institutions and platforms to develop validated monitoring tools rather than a direct claim that our current pipeline is ready for deployment. The manuscript already contains an explicit disclaimer that only surface-level diffusion is measured. Nevertheless, we accept that a dedicated error-analysis subsection is needed. We will add this subsection, discuss likely sources of false positives, and qualify the recommendations to stress that any operational monitoring must first demonstrate acceptable error rates on domain-specific validation data. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical screening with no derivations or self-referential reductions

full rationale

The paper performs an empirical screen of ~52,000 bioRxiv preprints using a hybrid lexical-LLM pipeline across fixed DURC/PEPP/governance categories drawn from external U.S. and Australia Group frameworks. No equations, fitted parameters, predictions derived from inputs, or self-citations appear in the load-bearing steps. The central claim rests on the pipeline's classifications of titles and abstracts, which are not defined in terms of the outputs themselves and do not reduce to any of the enumerated circularity patterns. The work is therefore self-contained as a descriptive mapping exercise.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The analysis depends on the untested assumption that the chosen DURC/PEPP categories and LLM judgments map to real biosecurity risk, plus the premise that surface text in titles and abstracts is a sufficient proxy for the presence of dual-use knowledge.

axioms (1)

domain assumption DURC, PEPP, and governance categories are aligned with U.S. and Australia Group oversight frameworks and can be applied to preprint metadata
Scoring is performed against these frameworks as stated in the abstract.

pith-pipeline@v0.9.1-grok · 5797 in / 1216 out tokens · 35239 ms · 2026-06-30T22:38:40.072500+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 13 canonical work pages

[1]

Chemberta-2: towards chemical foundation models.arXiv preprint arXiv:2209.01712,

Ahmad, W., Simon, E., Chithrananda, S., Grand, G., and Rosen, B. Chemberta-2: towards chemical foundation models.arXiv preprint arXiv:2209.01712,

work page arXiv
[2]

Common control list handbook, volume II: Biological weapons-related common control lists

Australia Group. Common control list handbook, volume II: Biological weapons-related common control lists. URL https://www.dfat.gov.au/sites/defaul t/files/australia- group- common- con trol-list-handbook-volume-ii.pdf . Ac- cessed: 2026-04-30. Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G. R., Wang, J., Cong, Q., Kinch, L. ...

2026
[3]

Bagal, V ., Aggarwal, R., Vinod, P., and Priyakumar, U

doi: 10.1126/science.abj8754. Bagal, V ., Aggarwal, R., Vinod, P., and Priyakumar, U. D. Molgpt: molecular generation using a transformer- decoder model.Journal of Chemical Information and Modeling, 62(9):2064–2076,

work page doi:10.1126/science.abj8754 2064
[4]

Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., Dafoe, A., Scharre, P., Zeitzoff, T., Filar, B., et al

doi: 10.1021/acs.jc im.1c00600. Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., Dafoe, A., Scharre, P., Zeitzoff, T., Filar, B., et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation.arXiv preprint arXiv:1802.07228,

work page doi:10.1021/acs.jc
[5]

Interpretable rna foundation model from unannotated data for highly accu- rate rna structure and function predictions.arXiv preprint arXiv:2204.00300,

Chen, J., Hu, Z., Sun, S., Tan, Q., Wang, Y ., Yu, Q., Zong, L., Hong, L., Xiao, J., King, I., et al. Interpretable rna foundation model from unannotated data for highly accu- rate rna structure and function predictions.arXiv preprint arXiv:2204.00300,

work page arXiv
[6]

K., McIlwraith, D., and Nair, S

Choudhari, J., Singh, P. K., McIlwraith, D., and Nair, S. Prompt smart, pay less: Cost-aware apo for real-world applications.arXiv preprint arXiv:2507.15884,

work page arXiv
[7]

Accessed: 2026-04-30

URL https://ou rworldindata.org/historical-pandemics . Accessed: 2026-04-30. Dauparas, J., Anishchenko, I., Bennett, N., Bai, H., Ragotte, R. J., Milles, L. F., Wicky, B. I., Courbet, A., de Haas, R. J., Bethel, N., et al. Robust deep learning–based pro- tein sequence design using proteinmpnn.Science, 378 (6615):49–56,

2026
[8]

Ferruz, N., Schmidt, S., and H¨ocker, B

doi: 10.1126/science.add2187. Ferruz, N., Schmidt, S., and H¨ocker, B. Protgpt2 is a deep unsupervised language model for protein design.Nature Communications, 13(1):4348,

work page doi:10.1126/science.add2187
[9]

Gracias, S., Le Seac’h, E., Donaire-Carpio, S., Vuillier, F., Vendramini, L., Moundib, A., Temmam, S., Rutkowska, M., Donati, F., Cupic, A., et al

doi: 10.1038/s414 67-022-32007-7. Gracias, S., Le Seac’h, E., Donaire-Carpio, S., Vuillier, F., Vendramini, L., Moundib, A., Temmam, S., Rutkowska, M., Donati, F., Cupic, A., et al. Entry, replication and innate immunity evasion of banal-236, a sars-cov- 2-related bat virus, in rhinolophus and human cells.PLoS pathogens, 22(4):e1013573,

work page doi:10.1038/s414
[10]

, year 2019

doi: 10.4135/9781071878781. URL https://methods.sagepub.com/book /mono/content-analysis-4e/toc. Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y ., et al. Language models of protein sequences at the scale of evo- lution enable accurate structure prediction.Science, 379 (6637):eade2574,

work page doi:10.4135/9781071878781
[11]

Evolutionary-scale prediction of atomic-level protein structure with a language model , volume =

doi: 10.1126/science.ade2574. Madani, A., Krause, B., Greene, E. R., Subramanian, S., Mohr, B. P., Holton, J. M., Olmos Jr, J. L., Xiong, C., Sun, Z. Z., Socher, R., et al. Large language models generate functional protein sequences across diverse fam- ilies.Nature Biotechnology, 41(8):1099–1106,

work page doi:10.1126/science.ade2574
[12]

doi: 10.1038/s41587-022-01618-2

doi: 10.1038/s41587-022-01618-2. Mak, K.-K., Wong, Y .-H., and Pichika, M. R. Artificial intelligence in drug discovery and development.Drug discovery and evaluation: safety and pharmacokinetic assays, pp. 1461–1498,

work page doi:10.1038/s41587-022-01618-2
[13]

Deep ignorance: Filtering pretraining data builds tamper-resistant safeguards into open-weight llms.arXiv preprint arXiv:2508.06601,

O’Brien, K., Casper, S., Anthony, Q., Korbak, T., Kirk, R., Davies, X., Mishra, I., Irving, G., Gal, Y ., and Biderman, S. Deep ignorance: Filtering pretraining data builds tamper-resistant safeguards into open-weight llms.arXiv preprint arXiv:2508.06601,

work page arXiv
[14]

Shuai, R

doi: 10.4049/jimm unol.1800708. Shuai, R. W., Ruffolo, J. A., and Gray, J. J. Generative language modeling for antibody design.Cell Systems, 13 (12):934–944,

work page doi:10.4049/jimm
[15]

United States Government

doi: 10.1016/j.cels.2021.11.003. United States Government. United states government pol- icy for oversight of dual use research of concern and pathogens with enhanced pandemic potential, May

work page doi:10.1016/j.cels.2021.11.003 2021
[16]

Accessed: 2026-04-30

URL https://worksinprogress.co/issue /pandemic-prevention-as-fire-fightin g/. Accessed: 2026-04-30. World Health Organization.Global guidance framework for the responsible use of the life sciences: mitigating biorisks and governing dual-use research. World Health Organization, Geneva,

2026
[17]

we enhanced aerosol transmissibility in ferrets

URL https://www.who.int/publicatio ns/b/65594 . Available under Creative Commons Attribution-NonCommercial-ShareAlike 3.0 IGO license. 7 The Biosecurity Blind Spot A. List of Keywords Pandemic Pathogens Influenza A virus, SARS-CoV-2, SARS-CoV-1, MERS-CoV , Monkeypox virus, Ebola virus, Marburg virus, Nipah virus, Hendra virus, Crimean-Congo hemorrhagic fe...

1918

[1] [1]

Chemberta-2: towards chemical foundation models.arXiv preprint arXiv:2209.01712,

Ahmad, W., Simon, E., Chithrananda, S., Grand, G., and Rosen, B. Chemberta-2: towards chemical foundation models.arXiv preprint arXiv:2209.01712,

work page arXiv

[2] [2]

Common control list handbook, volume II: Biological weapons-related common control lists

Australia Group. Common control list handbook, volume II: Biological weapons-related common control lists. URL https://www.dfat.gov.au/sites/defaul t/files/australia- group- common- con trol-list-handbook-volume-ii.pdf . Ac- cessed: 2026-04-30. Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G. R., Wang, J., Cong, Q., Kinch, L. ...

2026

[3] [3]

Bagal, V ., Aggarwal, R., Vinod, P., and Priyakumar, U

doi: 10.1126/science.abj8754. Bagal, V ., Aggarwal, R., Vinod, P., and Priyakumar, U. D. Molgpt: molecular generation using a transformer- decoder model.Journal of Chemical Information and Modeling, 62(9):2064–2076,

work page doi:10.1126/science.abj8754 2064

[4] [4]

Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., Dafoe, A., Scharre, P., Zeitzoff, T., Filar, B., et al

doi: 10.1021/acs.jc im.1c00600. Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., Dafoe, A., Scharre, P., Zeitzoff, T., Filar, B., et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation.arXiv preprint arXiv:1802.07228,

work page doi:10.1021/acs.jc

[5] [5]

Interpretable rna foundation model from unannotated data for highly accu- rate rna structure and function predictions.arXiv preprint arXiv:2204.00300,

Chen, J., Hu, Z., Sun, S., Tan, Q., Wang, Y ., Yu, Q., Zong, L., Hong, L., Xiao, J., King, I., et al. Interpretable rna foundation model from unannotated data for highly accu- rate rna structure and function predictions.arXiv preprint arXiv:2204.00300,

work page arXiv

[6] [6]

K., McIlwraith, D., and Nair, S

Choudhari, J., Singh, P. K., McIlwraith, D., and Nair, S. Prompt smart, pay less: Cost-aware apo for real-world applications.arXiv preprint arXiv:2507.15884,

work page arXiv

[7] [7]

Accessed: 2026-04-30

URL https://ou rworldindata.org/historical-pandemics . Accessed: 2026-04-30. Dauparas, J., Anishchenko, I., Bennett, N., Bai, H., Ragotte, R. J., Milles, L. F., Wicky, B. I., Courbet, A., de Haas, R. J., Bethel, N., et al. Robust deep learning–based pro- tein sequence design using proteinmpnn.Science, 378 (6615):49–56,

2026

[8] [8]

Ferruz, N., Schmidt, S., and H¨ocker, B

doi: 10.1126/science.add2187. Ferruz, N., Schmidt, S., and H¨ocker, B. Protgpt2 is a deep unsupervised language model for protein design.Nature Communications, 13(1):4348,

work page doi:10.1126/science.add2187

[9] [9]

Gracias, S., Le Seac’h, E., Donaire-Carpio, S., Vuillier, F., Vendramini, L., Moundib, A., Temmam, S., Rutkowska, M., Donati, F., Cupic, A., et al

doi: 10.1038/s414 67-022-32007-7. Gracias, S., Le Seac’h, E., Donaire-Carpio, S., Vuillier, F., Vendramini, L., Moundib, A., Temmam, S., Rutkowska, M., Donati, F., Cupic, A., et al. Entry, replication and innate immunity evasion of banal-236, a sars-cov- 2-related bat virus, in rhinolophus and human cells.PLoS pathogens, 22(4):e1013573,

work page doi:10.1038/s414

[10] [10]

, year 2019

doi: 10.4135/9781071878781. URL https://methods.sagepub.com/book /mono/content-analysis-4e/toc. Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y ., et al. Language models of protein sequences at the scale of evo- lution enable accurate structure prediction.Science, 379 (6637):eade2574,

work page doi:10.4135/9781071878781

[11] [11]

Evolutionary-scale prediction of atomic-level protein structure with a language model , volume =

doi: 10.1126/science.ade2574. Madani, A., Krause, B., Greene, E. R., Subramanian, S., Mohr, B. P., Holton, J. M., Olmos Jr, J. L., Xiong, C., Sun, Z. Z., Socher, R., et al. Large language models generate functional protein sequences across diverse fam- ilies.Nature Biotechnology, 41(8):1099–1106,

work page doi:10.1126/science.ade2574

[12] [12]

doi: 10.1038/s41587-022-01618-2

doi: 10.1038/s41587-022-01618-2. Mak, K.-K., Wong, Y .-H., and Pichika, M. R. Artificial intelligence in drug discovery and development.Drug discovery and evaluation: safety and pharmacokinetic assays, pp. 1461–1498,

work page doi:10.1038/s41587-022-01618-2

[13] [13]

Deep ignorance: Filtering pretraining data builds tamper-resistant safeguards into open-weight llms.arXiv preprint arXiv:2508.06601,

O’Brien, K., Casper, S., Anthony, Q., Korbak, T., Kirk, R., Davies, X., Mishra, I., Irving, G., Gal, Y ., and Biderman, S. Deep ignorance: Filtering pretraining data builds tamper-resistant safeguards into open-weight llms.arXiv preprint arXiv:2508.06601,

work page arXiv

[14] [14]

Shuai, R

doi: 10.4049/jimm unol.1800708. Shuai, R. W., Ruffolo, J. A., and Gray, J. J. Generative language modeling for antibody design.Cell Systems, 13 (12):934–944,

work page doi:10.4049/jimm

[15] [15]

United States Government

doi: 10.1016/j.cels.2021.11.003. United States Government. United states government pol- icy for oversight of dual use research of concern and pathogens with enhanced pandemic potential, May

work page doi:10.1016/j.cels.2021.11.003 2021

[16] [16]

Accessed: 2026-04-30

URL https://worksinprogress.co/issue /pandemic-prevention-as-fire-fightin g/. Accessed: 2026-04-30. World Health Organization.Global guidance framework for the responsible use of the life sciences: mitigating biorisks and governing dual-use research. World Health Organization, Geneva,

2026

[17] [17]

we enhanced aerosol transmissibility in ferrets

URL https://www.who.int/publicatio ns/b/65594 . Available under Creative Commons Attribution-NonCommercial-ShareAlike 3.0 IGO license. 7 The Biosecurity Blind Spot A. List of Keywords Pandemic Pathogens Influenza A virus, SARS-CoV-2, SARS-CoV-1, MERS-CoV , Monkeypox virus, Ebola virus, Marburg virus, Nipah virus, Hendra virus, Crimean-Congo hemorrhagic fe...

1918