Goodfire introduced monitoring probes for Baseten-hosted models, reporting lower test costs and detection of 94% of malicious hacking sessions.
Goodfire launched monitoring probes for AI models hosted by Baseten, giving developers a way to flag potentially unsafe agent behavior as a model runs. The company’s system examines internal model signals instead of relying only on a separate AI to review every output. The launch was reported on October 8, 2026.
Developers can select risks to watch for, including offensive hacking, misuse involving chemical or biological weapons, and reward hacking. They can also configure what happens after a warning: the event can be logged, sent to a person for review, or used to refuse a request. The probes can flag activity for further review by a separate AI model.
Goodfire tested its first monitor on Kimi K3. Across about 1,500 sessions, the company reported monitoring costs of roughly $51. TechCrunch reported comparisons of $233 for a less expensive AI model checking each step and about $10,000 for a top-tier model. In the same tests, Goodfire said the probes detected 94% of malicious hacking sessions and escalated 8.7% of harmless ones. The company also said running four probes added less than 2% to the time before the model began responding.
The approach uses intermediate calculations the model already produces, then applies classifiers to those signals. Goodfire CEO Eric Ho said this makes the monitoring cheaper than having another model reread the agent’s work. The figures are company-reported test results, not a guarantee of performance across other models or deployments.
The service is available to Baseten customers. Goodfire’s initial monitor targets Kimi K3, while the company’s longer-term research goal, as described by CTO Dan Balsam, is to trace model behavior back to where it emerged during training. Google DeepMind has also described deploying misuse-detection probes in Gemini, indicating that internal-signal monitoring is not unique to Goodfire.