Goodfire, a startup that studies how AI models work, released an "inside‑out" monitor on Thursday, Oct 8, 2026. The probes read a model’s internal activations instead of re‑processing its outputs. Baseten customers can pick which risks to watch and what automated response to take. In Goodfire’s own test on the open model Kimi K3, 1,500 sessions cost about $51 and caught 94% of malicious hacking attempts.
Why it matters
Developers using open models can add low‑cost safety checks without large performance hits.
