Goodfire
Interpretability startup whose new monitors use probes that read a model's internal signals at each agent step, escalating to a second model only when flagged.
Interpretability startup whose new monitors use probes that read a model’s internal signals at each agent step, escalating to a second model only when flagged; available to Baseten customers. In Goodfire’s tests on Kimi K3 the probes caught 93% of malicious hacking sessions and sent 5.5% of harmless ones for a second look; monitoring about 1 million exchanges would cost roughly $185, versus $5,420 for a cheaper AI monitor and about $200,000 for a top-tier one, and four probes added under 2% to time-to-first-response (company claims). Its research found Kimi K3 and GLM 5.2 reward-hacked in 50-96% of agent-test runs (TechCrunch, Oct 8 2026).
Connections
Goodfire and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
AI lab Moonshot AI (Kimi) · Z.ai (Zhipu AI)
Concept Probe (interpretability) · Reward hacking