Goodfire launches internal AI monitors for Baseten customers

Goodfire launched monitors for Baseten customers that inspect an AI model's internal signals rather than having another model review every step. Small detectors flag activity for closer inspection by a separate AI model. Customers can choose risks to monitor and responses including logging, human review or refusal. Goodfire's tests on Kimi K3 found that monitoring about 1,500 sessions cost roughly $51, versus $233 for a cheaper model checking every step. The probes caught 94% of malicious hacking sessions and referred 8.7% of harmless sessions for another look. Those results come from the company's tests.