Goodfire launched monitors for Baseten customers on October 8 that inspect a model’s internal activity as it works. The first monitors are built around Kimi K3, an open-weight model, and use small classifiers called probes to flag selected risks.
A separate model takes a closer look when a probe raises a flag. This avoids asking a larger reviewer to read every step of every agent session. Customers choose the risks they want monitored and the response, such as logging an event, escalating it to a person or refusing the request.
Goodfire reports that its probes caught 93% of malicious hacking sessions in its Kimi K3 tests, while 5.5% of harmless sessions were referred for further checking. Those referrals require a second check; a flag is not a confirmed misuse finding.
The company also reports lower monitoring costs because the probes reuse calculations the model is already performing.