arXiv subjects

Kemper Talley

Publications and source records attributed to Kemper Talley.

2 recordsLinked to original sources

arXiv2026

Building Barricades: Red-Teaming Agentic Biosynthesis Planning

The rapid democratization of biotechnology, enhanced by large language models (LLMs) and agentic software, is eroding longstanding barriers that have limited biosynthesis planning to experts. We present BioBulbul, an agentic system that couples LLMs with open-source biodesign tools for retrobiosynthesis, enzyme selection, gene cluster design, and DNA assembly planning. Using simple natural language prompts, BioBulbul autonomously generates plausible biosynthesis pathways and corresponding gene cluster sequences for several known small molecule toxins, demonstrating that non-experts could feasibly access them using an agentic system. Our red-teaming experiments confirm the technical viability of LLM-driven toxin biosynthesis while highlighting three principal barriers: scarcity of open-source biodesign tools, limited structured enzymatic reaction data, and the complexity of many small molecule toxin pathways. We argue that these barriers are likely to diminish within a decade, expanding the biological threat landscape. To mitigate emerging risks, we recommend expanding DNA synthesis screening to include chokepoint enzymes, extending chemical supplier controls to precursor compounds, establishing continuous AI-assisted threat surveillance pipelines, and developing robust LLM guardrails. Coordinated investment across government, industry, and AI developers is essential to safeguard against the misuse of agentic biosynthesis planning.

q-bio.OT↗

arXiv2024

The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning

The White House Executive Order on Artificial Intelligence highlights the risks of large language models (LLMs) empowering malicious actors in developing biological, cyber, and chemical weapons. To measure these risks of malicious use, government institutions and major AI labs are developing evaluations for hazardous capabilities in LLMs. However, current evaluations are private, preventing further research into mitigating risk. Furthermore, they focus on only a few, highly specific pathways for malicious use. To fill these gaps, we publicly release the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of 3,668 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. WMDP was developed by a consortium of academics and technical consultants, and was stringently filtered to eliminate sensitive information prior to public release. WMDP serves two roles: first, as an evaluation for hazardous knowledge in LLMs, and second, as a benchmark for unlearning methods to remove such hazardous knowledge. To guide progress on unlearning, we develop RMU, a state-of-the-art unlearning method based on controlling model representations. RMU reduces model performance on WMDP while maintaining general capabilities in areas such as biology and computer science, suggesting that unlearning may be a concrete path towards reducing malicious use from LLMs. We release our benchmark and code publicly at https://wmdp.ai

cs.LG↗