AI and the Potential for Biological Threats

A new report suggests that artificial intelligence might be capable of designing biological threats that are even more lethal than COVID-19.

Published by RAND on Tuesday, the report explores various strategies to prevent malicious entities from misusing AI for the creation of biological weapons. This comes in light of announcements from Anthropic, which revealed that its AI system, Claude, successfully designed proteins targeting 14 out of 15 objectives in a laboratory environment.

Previously, leaders in AI from companies like Google, OpenAI, and Anthropic urged Congress to mandate that sellers of DNA and RNA implement buyer screenings and maintain accurate records of transactions. This plea was reported by the Wall Street Journal in June.

The report expresses concern that “AI-enabled biotechnology might lower the barriers to creating biological weapons, in terms of technical, operational, and motivational aspects.” It warns that such advancements could facilitate catastrophic attacks involving new or enhanced pathogens that might rival or exceed the death toll of past pandemics.

Furthermore, the findings underline the possibility of disasters that could be significantly more severe than COVID-19. The report’s Purpose section states, “[This report] focuses on high-consequence AI-enabled biological incidents, which we define as biological events capable of producing impacts on the order of ten times the scale of the coronavirus disease 2019 (COVID-19) pandemic or greater, societal disruption, or systemic destabilization.”

Both RAND and Anthropic have yet to respond to requests for comments regarding these findings.

Anthropic has recognized the inherent dangers of this research and is taking steps to limit access to these AI capabilities for the public. Their announcement highlighted, “Without robust safety measures, these technologies could enable bad actors to conduct dangerous research, like developing bioweapons.” As part of their commitment to safety, certain capabilities, like protein design, are currently restricted in Claude Fable 5.

RAND’s report emphasizes the complexity of managing such risks. It points out that, “No single safeguard can prevent a bad actor from using AI to help build a biological weapon.” However, they also identify nine different interventions that, when combined, could significantly reduce the chances of a major AI-enabled attack.

Steph Guerra, the lead author of the report, expressed concerns that “the most dangerous thresholds have not yet been crossed.” She maintained that although this unregulated technology poses a genuine threat, it can be managed with “mutually reinforcing approaches.”