Steven T. Murphy

This study evaluates the effectiveness of biosecurity screening tools against AI-generated synthetic homologs of proteins of concern, finding some tools can detect fragments as short as 50 nucleotides.

The paper develops a rubric to define sequences of concern by analyzing test sequences and refining categorization through stakeholder review, reducing disputed categorization by 44.3% for controlled viruses.

The paper evaluates sequence screening tools using a NIST dataset to assess their consistency in identifying sequences of concern.

The paper evaluates vulnerabilities in biosecurity screening tools by testing their ability to detect splitting-based obfuscation of synthetic DNA sequences.

The paper evaluates biosecurity screening tools for detecting AI-generated synthetic protein homologs and reports improved detection rates through patches.