AI interpretability startup Goodfire launched a new AI monitor on Thursday, designed to capture anomalous AI agents at a lower cost, TechCrunch reported. The system monitors the internal workings of AI models directly, rather than their external outputs, and only calls in a secondary AI for deeper inspection when anomalies are detected. Goodfire stated that this "inside-out" approach is more cost-effective than traditional methods. Tests on the Kimi K3 model showed that monitoring 1,500 sessions cost approximately $51, significantly less than the $233 to $10,000 range for other solutions. The system can capture 94% of malicious hacking sessions and sends only 8.7% of harmless sessions for review. Goodfire's monitor is now available to Baseten customers and can be used to detect risks such as offensive hacking, chemical and biological weapon misuse, and reward fraud.