Expanding Our Analysis Of Biological Ai Models

2026-07-14 · Source: Papers & Reports | Epoch AI · Field: Science & Research — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Research Methodology & Innovation · Depth: Intermediate, extended

Summary

This report presents an expanded database of AI models in biology, commissioned by Sentinel Bio, building on a 2024 collaboration. The database comprises 1,196 models, with 1,124 annotated via AI assistance and 72 manually reviewed. Key findings reveal that pre-release risk assessments (2.5%) and risk-related evaluations (2.3%) are rare, though more common among notable models, primarily driven by frontier LLMs. Documented safeguards are also scarce (3.2% overall), with frontier LLMs accounting for over half of those. Safeguard types vary, with LLMs favoring inference-time filtering, while non-LLM biological AI models more commonly use post-training safety techniques or training data filtering. Most models share inference code (58%) and training data (46%), but open weights are less frequent (23%). Protein engineering and small biomolecule design constitute the largest model categories, and approximately one in five models (253 of 1,196) are finetuned from existing foundation models, with ESM-2 being the most prevalent base.

Key takeaway

For AI Scientists developing biological models, you should prioritize integrating robust pre-release risk assessments and evaluations into your development lifecycle. The current low adoption rates (2.5% for assessments) indicate a critical gap, especially for non-LLM biological AI. Consider implementing post-training safety techniques and training data curation, as these are more common and effective for biology-specific models than inference-time filtering. This proactive approach is crucial for mitigating biosecurity risks and fostering responsible innovation.

Key insights

Biological AI models largely lack documented risk assessments and safeguards, except for frontier LLMs.

Principles

• Risk assessments and safeguards are rare in biological AI.
• Frontier LLMs drive most documented safety practices.
• Open weights are less common than open code/data.

Method

The database was built by searching academic databases, filtering candidates with language models, extracting metadata, and manually reviewing notable models.

In practice

• Implement pre-release risk assessments and evaluations.
• Adopt post-training safety techniques for non-LLM biological AI.
• Consider training data curation for biosecurity.

Topics

• Biological AI Models
• Biosecurity
• Risk Assessment
• AI Safeguards
• Protein Engineering
• Foundation Models (ESM-2)

Best for: Research Scientist, AI Scientist, AI Ethicist

Related on AIssential

• Biological Sequence Models In The Context Of The Ai Directives — The rapid scaling of biological sequence models and data necessitates updated AI governance to address emerging dual-use risks.
• Muse Spark Safety & Preparedness Report — Meta's Muse Spark LLM was deemed safe for deployment after mitigating "high risk" chemical and biological capabilities.
• Bioinfohazards: Jassi Pannu on Controlling Dangerous Data from which AI Models Learn — AI advancements heighten engineered pandemic risks, necessitating data controls and a multi-layered biosecurity defense strategy.
• Measuring Biological Capabilities and Risks of AI Agents — Interpreting AI agent biological risk evaluations critically depends on transparent design, execution, and documentation choices.
• The Once And Future Fable #5 — The "Mythos Moment" reveals an ad hoc, reactive US AI policy struggling with frontier model control, legal frameworks, and open-source risks.
• Open Thread 428 — AI safety and biosecurity funding opportunities are expanding, alongside community updates and niche service advertisements.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Papers & Reports | Epoch AI.