Goodfire AI Launches Cost-Effective Monitoring System for Hidden Model Reasoning

Goodfire, an AI safety startup, has unveiled monitors that detect harmful behavior in reasoning models without requiring access to explicit reasoning traces—a breakthrough for systems that perform computation internally before generating outputs. The system, launched on October 8, 2026, on the Baseten inference platform, uses small detectors called probes to examine a model's internal neural activations at each step of its reasoning process, catching misalignment signals that would be invisible if monitors relied solely on final outputs.