
Goodfire deploys production cyber monitors on Kimi K3 and GLM 5.3
An 8 October research post says probe-plus-judge cascades on Kimi K3 and GLM 5.3 hit 93% recall at a 5.5% false-positive rate and cost about 50 times less than an always-on LLM judge. Those figures are Goodfire’s. FAR.AI’s quoted check used a two-day static jailbreak battery in chat, not an adaptive agent red team.















