Open-Model Safety Monitoring Gets 50x Cheaper as Goodfire Deploys Probes for Kimi K3 and GLM 5.3
Goodfire published a research post on October 8, 2026 detailing probe-based cybersecurity monitors deployed on a production inference stack for the open models Kimi K3 and GLM 5.3. The monitors match an LLM judge's recall at Goodfire's operating point while costing roughly 50 times less. This means production safety monitoring for open-weight models can shift from expensive LLM-judge calls to lightweight probes without sacrificing detection coverage. The deployment targets two specific open models, suggesting the approach is tuned to their internal representations rather than being a generic API wrapper.
breakthroughs10/8/2026