Goodfire deploys internal-signal probe monitors to catch risky AI agent behavior

Goodfire Research has built and deployed probe-based cyber monitors for Kimi K3 and GLM 5.3 that read a model's internal signals rather than only its written output.

A lightweight probe screens suspicious exchanges before an LLM judge reviews them. Goodfire says this cascade reaches about 93% recall with a 5.5% benign-session interruption rate, at roughly 50x lower judge cost. In its tests on Kimi K3, monitoring about 1,500 sessions cost roughly $51 versus about $10,000 for a top-tier AI judge, catching 94% of malicious hacking sessions. These figures are Goodfire's own claims and have not been independently verified.

The company cites red-teaming by FAR.AI, its one independent element, which Goodfire says reduced universal jailbreaks to zero across 140 strategies and cut total jailbroken interactions by 97%. Goodfire describes these as results from a static battery of attacks.

TechCrunch reports the monitors are now available to customers of Baseten, a company that hosts and runs AI models for others. Customers can choose which risks to watch, including offensive hacking, chemical and biological weapons misuse, and reward hacking.