Goodfire unveils "inside-out" monitors for rogue SI agents — cheaper than checking their outputs
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
SI agents can be watched more cheaply by looking inside the model rather than just reading its answers; how this performs outside the company's own tests is not yet clear.
According to TechCrunch, Goodfire's monitors read a model's internal signals with small "probes" and send anything suspicious to a second SI system for review; the response can be logging, human review or denying the request. The company says monitoring about 1,500 sessions cost roughly $51, versus $233 for cheaper alternatives and $10,000 for pricier ones; four probes add less than 2% latency. In tests with the Kimi K3 model, per Goodfire, 94% of malicious hacking sessions were caught, with 8.7% of harmless ones sent for review. The monitors are available through the Baseten platform.
Sources 1