Goodfire launches cheaper inside-out AI agent monitors

Goodfire, an AI interpretability startup, has launched monitors that watch an AI model's internal signals rather than rereading its outputs, cutting costs sharply. In tests on the open model Kimi K3, monitoring one million exchanges cost roughly $185, versus $5,420 for a cheaper external AI checker. The probes caught 93% of malicious hacking sessions and added under 2% latency. Available to Baseten customers, the monitors target risks including offensive hacking and reward hacking, and let operators choose automated responses. Goodfire says the approach is especially suited to open models, which lack built-in safeguards.