“Even after being patched with details of AI-generated bioweapons, none of the tested biosecurity screening tools could identify all proteins with high in-silico harm metrics.”