Goodfire launches Baseten tool inspecting model internals and flagging risks
Technology
Goodfire just dropped an AI monitoring system that checks what's happening inside AI models, not just their outputs.
Available for Baseten users, it acts like airport security for AI, using probes to spot risks and automatically block or flag suspicious activity.
Kimi K3 testing $51 detected 94%
This system is a game changer for cost: running 1,500 sessions on the Kimi K3 model was just $51 (compared with $233 for a cheaper AI model checking every step).
It caught 94% of bad sessions but did flag about 8.7% of safe ones by mistake.
Plus, even when using multiple probes at once, response times barely slow down; great news for anyone building open models that need extra protection.