Hidden Demons - Engineering For The Unknown

Stability Engineering – Volume 9 www.manhattanproject20.com 1 117-263-D Hidden Demons Are Unknown Latent Behaviors Engineering for the Unknown-Unknowns of Artificial Intelligence 1. The Hidden Demon • As AI systems become increasingly capable, agentic and deeply coupled with the world, we are encountering a difficult engineering reality. • A model may behave acceptably throughout training, evaluation and red teaming—and later exhibit behavior its designers did not anticipate.