DeepMind introduces an AI biosecurity initiative.

      Google DeepMind has initiated a program aimed at addressing biological threats through AI. The London-based lab, along with its affiliate Isomorphic Labs, refers to this initiative as bioresilience, as outlined in a joint blog post.

      The objectives are twofold. DeepMind aims to prevent the misuse of its models while assisting governments, scientists, and biosecurity organizations in utilizing AI for the prevention, detection, and response to outbreaks. Over the past year, it reports having established more than 15 partnerships in this area.

      Program Details

      For prevention, DeepMind implements a four-step safety protocol: threat modeling, evaluations, mitigations, and monitoring. Additionally, it is modifying its SynthID watermarking for biological applications, enabling DNA-synthesis providers to screen for potentially dangerous AI-generated sequences.

      Regarding detection, the company is employing its AlphaEvolve agent to enhance the algorithms involved in pathogen sequencing, which it claims will expedite and reduce the cost of tracking new outbreaks.

      In terms of response, DeepMind is providing trusted researchers with access to its latest models to aid in the design of vaccines and other countermeasures. Isomorphic Labs has established a unit dedicated to deploying its drug design engine in the event of outbreaks.

      Access with Limitations

      DeepMind is expanding access to certain models and agents for trusted governments, researchers, and biosecurity partners. These are classified as “low-risk” restricted releases, rather than public launches, as reported by Axios.

      The company states it has corresponding controls in place. “If… we were to determine that we had reached a critical capability level without the appropriate mitigations, we would not proceed with a launch,” said Helen King, DeepMind’s vice president of responsibility, in a statement to Axios. She indicated that this threshold has not yet been reached.

      Part of a Broader Safety Initiative

      This program comes shortly after DeepMind's chief, Demis Hassabis, called for governments to create a standards body for frontier AI. Competing labs are currently discussing similar safety measures, including testing protocols and a potential pause on frontier AI. This efforts fits within DeepMind’s broader focus on mitigating chemical, biological, radiological, and nuclear risks.

      For years, Google has been at the intersection of AI and biology. Its AlphaFold tool for protein structure determination has earned Hassabis and John Jumper a portion of the 2024 Nobel Prize in Chemistry, while Isomorphic Labs is now commercializing that technology for drug discovery.