Goodfire built real-time cyber monitors for Kimi K3 and GLM 5.3

What is changing?

Goodfire says it deployed probe-based cybersecurity monitors for Kimi K3 and GLM 5.3 on a production inference stack. The system uses small classifiers that watch internal model activations on every token as a first-line filter, sending only suspicious interactions to an LLM judge.

According to the company's October 8 technical post, the cascade reached 93% recall at a 5.5% false-positive rate at its chosen operating point. Goodfire reports a cost below $200 per million exchanges, roughly 50 times less than judging every turn with an LLM, with no measurable inference latency.

In end-to-end session tests, the monitor raised Kimi K3's harmful-session blocking rate from about 29% to 96%, and GLM 5.3's from about 36% to 91%. Harmless-session interruption also rose, from roughly 5% to 9% and from 10% to 15%, respectively. These figures come from the provider's own evaluations.

The main advantage is speed: risky tool calls can be stopped before execution. The thresholds still require a trade-off between false positives and safety, and broader independent testing is needed before generalizing the results.