CBRN content filtering and frontier safety evaluations

Control objective

This control checks that a generative summarization-and-chat system does not produce or amplify dangerous chemical, biological, radiological, or nuclear (CBRN) content, and that the underlying model is periodically tested for newly emerging hazardous capabilities. It matters because frontier models can inadvertently surface weaponization-relevant information, a severe safety and reputational risk that a bank deploying public-facing or internal assistants must actively contain. No metric or formula is specified, so the control is governed by its threshold rather than a computed score. Implement it by running a CBRN classifier that screens content across the four hazard categories (C/B/R/N) on the model's inputs and outputs, wiring it inline so flagged content is blocked or escalated before reaching the user, and logging each classification decision and triggered block as evidence; separately, schedule a structured frontier capability evaluation each quarter to probe whether the model has gained dangerous abilities, retaining the eval reports. The threshold requires the CBRN classifier to run on 100% of relevant content and a frontier capability evaluation to be conducted quarterly; any relevant content that bypasses the classifier, or a missed quarterly evaluation, is a breach that must trigger immediate remediation of the filtering coverage or the evaluation cadence.