Goodfire releases internal activation monitors for AI agents on October 8
Goodfire internal probes reuse forward pass computations and detected 94 percent of malicious hacking sessions at a monitoring cost of about $51 for roughly 1,500 sessions.
Interpretability company Goodfire announced internal activation probes for monitoring AI agents on October 8 and made them available to customers through Baseten. Instead of using an external model to read every output, the probes inspect intermediate activations already computed by the main model and run small classifiers over them. A second model performs deeper review only after a probe raises a flag. Goodfire said monitoring about 1,500 sessions cost roughly $51, compared with about $233 for a cheaper external model checking every step and about $10,000 for a premium model. In tests based on Kimi K3, the system detected 94 percent of malicious hacking sessions and sent 8.7 percent of harmless sessions for secondary review. Running four probes increased initial response time by less than two percent. Customers can configure monitored risks and responses including logging, human review or request rejection.