Goodfire says its new ‘inside-out’ monitors catch rogue AI agents at a fraction of the cost
By Pradhyuman,
Goodfire launched AI monitors that track internal signals inside a model while it runs. The monitors are available to Baseten customers. Standard monitoring uses a second AI model to read output text. Goodfire detectors tap into the model's existing calculations during execution.
Goodfire CEO Eric Ho said internal activation monitors reuse forward pass computations. In tests on the Kimi K3 model, 1 million exchanges cost $185 to check with Goodfire probes. Checking the same exchanges cost $5,420 with a smaller secondary model and $200,000 with a top-tier model.
Goodfire stated its probes caught 93% of malicious hacking sessions and added less than 2% to response start times. Goodfire CTO Dan Balsam said the system helps users detect potential hacking risks before they occur.
Maintained by Pradhyuman.
Filed under: Models & Research, Policy & Safety