Biosecurity monitors claimed to catch more unsafe requests than frontier-model safeguards

GoodfireAI says it built biosecurity monitors for AI agents working on biology tasks. The company claims the monitors use protein model embeddings to screen requests in milliseconds, catch more unsafe biological requests and produce fewer false positives than frontier-model safeguards, while refusing fewer dual-use tasks. It also says its method is three to five times more robust against adversarial attacks than well-established screening methods.