SecureBio

Biosecurity beyond model benchmarks

SecureBio is a US charitable nonprofit working on prevention, detection and defense against catastrophic pandemics. Its AI team evaluates whether models could increase biological misuse capabilities and researches ways to reduce those risks. This sits alongside work on biological detection rather than replacing the organization's broader biosecurity mission. [1]

A documented role in frontier evaluations

SecureBio developed the Virology Capabilities Test and other targeted biology evaluations. Its June 2025 overview identifies assessments of models from several developers and explains that much of the work remains confidential. Published system cards provide a partial record of those contributions. [2]

The Claude 4 system card credits SecureBio with evaluation tasks and identifies its VCT benchmark. This is evidence of work used to assess the release, not a claim that SecureBio trained Claude or approved every deployment decision. [3]

What the tests establish

A targeted test measures specified knowledge or task performance under its evaluation conditions. SecureBio describes several complementary methods because biological risk cannot be reduced to a single score. Readers need the particular test, access conditions and assessment scope when interpreting a model developer's safety conclusion. [2][3]