Safety3d ago
Safety
LLMs Struggle With Unclassified Biological Threats
Mirror RNA polymerase creation highlights a critical gap in AI safety filters. Current models often rely on official CBRN categories to trigger refusals, leaving them blind to emerging risks like mirror life. This research suggests that LLMs fail to identify novel threats until regulators formally classify them. Practitioners must develop more proactive, heuristic-based safety guardrails.