Goodfire Launches Low‑Cost Inside‑Out Monitors to Detect Rogue AI Agents

Goodfire has introduced a new class of safety monitors that observe the internal activations of an AI model while it runs, rather than relying on a separate model to re‑read its output.

Why it matters for GPU / AI infrastructure

By tapping into computations already performed during the forward pass, the probes add virtually no latency and keep operating costs low, making continuous oversight feasible for large‑scale inference workloads.

Baseten customers can select specific threat vectors—such as illicit hacking attempts, chemical‑biological misuse, or reward hacking—and define automated responses ranging from logging to outright request rejection.

In early tests on the open Kimi K3 model, the system detected 94 % of malicious sessions for roughly fifty dollars, compared with hundreds or thousands of dollars for conventional monitoring approaches.