Definition
This refers to technical restrictions meant to stop AI agents from being used to design dangerous biological agents, by blocking access to specialized biology design tools. New research shows these software-level restrictions can often be bypassed, meaning they don't provide the safety guarantee many assumed.
Why it matters
Policymakers and lab safety teams relying on 'we blocked access to the dangerous tools' as their main safeguard need to know that assumption has now been empirically challenged.