Goodfire Announces New System to Monitor AI Agents
Goodfire has developed new monitors that track the internal signals of AI models, enabling the detection of malicious agents at a lower cost.
AI startup Goodfire has launched more economical inside-out monitoring systems that observe the internal signals of models in real-time during operation.
Inside-Out Monitoring Method
Traditionally, supervising AI agents involved examining the outputs of a second AI model. However, this method rapidly increased costs during long-running operations.
The new system developed by Goodfire offers an alternative approach by directly monitoring the processes running inside the AI model rather than just reading the text generated.
System Similar to Airport Security
The working logic of the system is compared to airport security screenings. Signal detectors scan internal signals at every step of the agents' operations to detect suspicious situations.
Only when a suspicious situation is flagged by the detectors does a separate AI model step in to conduct a detailed review.
Cost Advantage and Test Results
Because the new system reuses the calculations the model is already performing, it significantly reduces the cost and time of running a separate model.
Tests indicated that the system reduces costs to much lower levels compared to traditional methods and successfully catches a large portion of cyber attack sessions.