The company says probes reduced monitoring costs in tests on Kimi K3 while detecting 94% of malicious hacking sessions.
Goodfire launched monitors that inspect AI models’ internal signals through Baseten on October 8, TechCrunch reported. The company says the approach costs less than having another model continuously review an agent’s output.
Small detectors called probes inspect internal activations during operation. A separate model reviews activity flagged by those probes. Customers can monitor hacking, weapons misuse and reward hacking, then choose logging, human review or request refusal.
In company tests covering about 1,500 Kimi K3 sessions, monitoring cost approximately $51, compared with $233 for a lower-cost model reviewing every step and roughly $10,000 for a higher-end model.
Goodfire said its probes detected 94% of malicious hacking sessions and referred 8.7% of benign sessions for additional review. Running four probes increased initial response latency by less than 2%, according to the company.
The service follows a safety partnership announced last month by Baseten’s Base Labs, Goodfire and Hugging Face.
