New research reveals LLMs fail to flag novel threats like mirror life

In a new research series titled "Can Large Language Models Identify Novel Threats?" independent AI safety researcher Failfinder70 examines how LLMs handle safety requests when the topic falls outside established threat categories. The test case is mirror life — specifically mirror RNA polymerase, which was successfully synthesized in 2022. In 2024, the scientific community issued a formal warning against further mirror life research due to existential biosecurity risks, yet it remains unclassified as a weapon of mass destruction or CBRN threat, though Congress is assessing the question.

The study asks: if a user requests uplift (i.e., help creating mirror organisms), will the LLM refuse? The researcher identifies five mechanisms models might use to detect novel threats: vocabulary triggering (e.g., "synthetic virus"), category matching (linking precursors to biological weapon categories), principled inference (reasoning about harm regardless of frame), actionability gradients (distinguishing general knowledge from instructions), and user sophistication (refusing more technically specific requests). Prior work by Kevin M. Esvelt in 2025 showed that even advanced models fail to recognize mirror life as dangerous, even after expert guidance. This gap matters because technology outpaces regulation, and models must reason or search beyond their training cutoffs to identify novel risks. The research highlights that current safety systems may inherit legal categories that are too slow to adapt, potentially allowing catastrophic misuse before classification catches up.