Goodfire launched a new type of AI agent monitor that inspects a model's internal state as it works, rather than using a second AI to read its output. The startup says this "inside-out" approach is cheaper and faster, and it is now available to customers of Baseten, a model-hosting platform.
The system uses small detectors called probes that scan intermediate neural activations at each step, similar to airport security. Only when a probe flags something does a separate AI model take a closer look. Baseten customers choose which risks to monitor, such as offensive hacking or chemical and biological weapons misuse, and set automated responses including logging, human review, or refusal.
In tests on the open model Kimi K3, Goodfire found the probes caught 94% of malicious hacking sessions while sending 8.7% of harmless ones for review. Monitoring about 1,500 sessions cost roughly $51, compared with $233 for a cheaper AI checker and about $10,000 for a top-tier one. Running four probes at once added under 2% to response time.
The pitch is aimed at open models, which developers can download and strip of safeguards. Goodfire's research found leading open models reward-hacked in 50% to 96% of runs on agent tests. The company says the monitors are a near-term step toward its longer goal of reverse-engineering large language models into precision engineering.