Researchers assessed biosafety risks when using AI

A group of scientists published a comprehensive review dedicated to the impact of generative models on biological research and potential security threats. The work analyzes how LLM and specialized biological models can be used for planning experiments, generating protein sequences, and automating laboratory cycles.

The authors note that current state-of-the-art systems already surpass experts in in-silico tasks and methods for bypassing screening systems. However, the physical implementation of biological threats remains a complex task due to the lack of "implicit knowledge" and limitations when working in real "wet" laboratories. The study emphasizes that the risk depends not only on the capabilities of AI, but also on the user's intentions, access to materials, and the presence of protective mechanisms.

The article criticizes the current approach to security: alignment methods, effective for general models, poorly adapt to biological systems. Researchers propose implementing a defense-in-depth strategy, which links the thresholds of model capabilities with proportional responsibility of developers. This should help reduce the risks of creating dangerous pathogens while preserving benefits for medicine and biotechnology.

Particular attention is paid to the role of model interpretability in auditing hazardous skills. The authors insist on the need to create tools capable of verifying the absence of malicious functions in biological foundation models. The work is distributed under the arXiv license and sets a vector for future regulatory standards in the field of AI biotechnology.