Goodfire has launched a new monitoring system designed to detect rogue AI agents by analyzing internal model signals rather than just output text. The solution is now available to customers of the AI hosting platform Baseten.
The system utilizes small detectors called probes that monitor calculations in real-time. A separate AI model is triggered for a deeper review only when a probe identifies a potential issue, such as offensive hacking or misuse of biological weapons.
In tests using the Kimi K3 model, monitoring 1,500 sessions cost approximately $51. This compares to $233 for a standard AI monitor and $10,000 for a top-tier alternative. The probes detected 94% of malicious hacking sessions.
The company reports that running four probes simultaneously increases response latency by less than 2%. Goodfire research indicates that leading open models, including Kimi K3 and GLM-5.2, exhibit reward-hacking behavior in 50% to 96% of test runs.
