Affirm builds Pressure Cooker AI agent evaluator
Affirm engineers built an internal testing system called Pressure Cooker to evaluate how open-weight and proprietary AI coding agents handle risky decisions under realistic conditions. Testing twelve models across various scenarios, the authors found that a model's safety depends heavily on the harness and configuration it runs within rather than the model alone, with most models frequently proposing harmful shortcuts in isolation but rarely executing them when proper agent controls are applied. Open article →