Forecasting LLM‐Enabled Biorisk and the Efficacy of Safeguards
Risk Analysis · 2026-11 · 46 (11)
DOI: 10.1111/risa.70397 ↗At a glance
Journal article in Risk Analysis, published online on 2026-10-28. Written by Bridget Williams and 12 co-authors. Institution: Forecasting Research Institute, Claymont, Delaware, United States. Major MEDLINE topics: Large Language Models, Security Measures. 39 references. Funded by: Open Philanthropy., Open Philanthropy Project.
Abstract (original language)
Capabilities of large language models (LLMs) on several biological benchmarks have prompted not only excitement about their usefulness for research but also concern about biosecurity risks. We recruited 46 subject‐matter experts in biology and biosecurity, and 22 generalist forecasters to estimate the risks of growing LLM capabilities. Most respondents expected some hypothetical LLM capabilities to increase their estimates of risk and dramatically underestimated the rate of LLM progress in those capabilities. When these capabilities arose, the median forecast did update as anticipated. The median expert initially predicted a 0.3% annual risk of a human‐caused epidemic that causes 100,000 deaths. This estimate then rose to 1.5% conditional on several hypothetical LLM capabilities, including matching the performance of a top‐performing team of virologists on a virology troubleshooting test, which the median expert did not expect to occur until 2030. Given this finding, we conducted a baselining study and found that LLMs already crossed this performance threshold in April 2025. A follow‐up survey conducted after these results were available found that forecasters significantly updated their median unconditional risk estimates upward, with the new median expert forecast of annual risk rising to 1.5% (as anticipated by the earlier conditional forecasts). More encouragingly, most experts thought that the adoption of LLM safeguards and mandatory nucleic acid screening would negate the risk associated with some hypothetical LLM capabilities.
Subjects
MeSH terms (MEDLINE)
Author keywords
Authors and institutions
- Forecasting Research Institute, Claymont, Delaware, United States.
- Centre for the Governance of AI, London, United Kingdom.
- SecureBio, Cambridge, Massachusetts, United States.
- Johns Hopkins Center for Health Security, Baltimore, Maryland, United States.
- Federal Reserve Bank of Chicago, Chicago, Illinois, United States.
- Wharton School of the University of Pennsylvania, Philadelphia, Pennsylvania, United States.
ORCID profiles: Bridget Williams, Rebecca Ceppas de Castro, Otto Kuusela
Funding
- Open Philanthropy.
- Open Philanthropy Project
Cite
APA
Williams, B., Righetti, L., Rosenberg, J., de Castro, R. C., Kuusela, O., Britt, R., Soice, E., Morales, A., Sanders, J., Donoughe, S., Black, J., Karger, E., & Tetlock, P. E. (2026). Forecasting LLM‐Enabled Biorisk and the Efficacy of Safeguards. Risk Analysis, 46(11). https://doi.org/10.1111/risa.70397
Vancouver
Williams B, Righetti L, Rosenberg J, de Castro RC, Kuusela O, Britt R, et al. Forecasting LLM‐Enabled Biorisk and the Efficacy of Safeguards. Risk Analysis. 2026;46(11). doi:10.1111/risa.70397
BibTeX
@article{Williams2026,
author = {Williams, Bridget and Righetti, Luca and Rosenberg, Josh and de Castro, Rebecca Ceppas and Kuusela, Otto and Britt, Rhiannon and Soice, Emily and Morales, Alvaro and Sanders, Jon and Donoughe, Seth and Black, James and Karger, Ezra and Tetlock, Philip E},
title = {Forecasting LLM‐Enabled Biorisk and the Efficacy of Safeguards},
journal = {Risk Analysis},
year = {2026},
volume = {46},
number = {11},
doi = {10.1111/risa.70397}
}
EUVIMED