Startup Goodfire Launches Internal Monitoring Probes to Catch Rogue AI Agents at Lower Costs

Startup Goodfire launched a new internal monitoring system on Thursday designed to catch rogue AI agents at a fraction of the cost of traditional oversight methods. Instead of using a second AI model to read everything an agent writes, Goodfire employs small detectors called probes. These probes read the model's internal neural signals at every step, acting like a security scanner.

A separate AI model only takes a closer look if a probe flags a potential risk. Baseten customers can choose which risks to track, such as hacking or weapons misuse, and decide whether to log events, request human review, or block requests. According to Goodfire, the approach is much cheaper because the probes reuse calculations the model is already making.

In company tests monitoring Kimi K3, 1,500 sessions cost about $51. By comparison, a cheaper reviewing model cost $233, while a top-tier model cost roughly $10,000. Running four probes added less than two percent to the model's response time, and the system caught 94 percent of malicious hacking sessions. Goodfire built its first monitor around the open Kimi K3 model.

The company aims its product at open models, which developers can strip of built-in safeguards.