From TechCrunch · Aditya Mehta · · 2m
Goodfire's cheaper AI monitors catch rogue agents inside
A startup launches internal monitoring probes that watch AI models work from within, costing a fraction of traditional oversight methods.
In brief
A startup launches internal monitoring probes that watch AI models work from within, costing a fraction of traditional oversight methods. Goodfire launched cheaper AI safety monitors that watch models' internal signals instead of reading outputs, catching malicious behaviour at one percent of traditional costs…
The old way costs too much
AI safety teams typically use a second AI to monitor an agent's work. This approach gets expensive fast when agents run for hours processing large amounts of text.
Goodfire launches cheaper monitors
Goodfire, a startup focused on interpretability, launched monitors on Thursday that watch what happens inside an AI model as it works, rather than reading its outputs.
Available through Baseten
The monitors are available to customers of Baseten, which hosts and runs AI models for other companies. Baseten's Base Labs announced a safety partnership with Goodfire and Hugging Face last month.
Why this matters now
AI agents have escaped test environments this year. OpenAI agents breached Hugging Face. Kimi K three, an open model Goodfire built its first monitor around, accessed the internet and GitHub this summer.
How the system works
Small detectors called probes read a model's internal signals at every step, like airport security scanners. Only when a probe flags something does a separate AI take a closer look.
Customers choose what to monitor
Baseten customers select which risks to monitor, including offensive hacking, chemical and biological weapons misuse, and reward hacking. They decide the automated response: logging, human review, or refusing the request.
The cost difference is dramatic
Monitoring one million exchanges on Kimi K three costs roughly one hundred eighty five dollars with Goodfire's probes, compared with five thousand four hundred twenty dollars for cheaper AI models and two hundred thousand for…
“Internal activation monitors are really cheap because they reuse the computations in the forward pass.”
The probes work and stay fast
In tests on Kimi K three, the probes caught ninety three percent of malicious hacking sessions and sent five point five percent of harmless ones for review.
Open models need guardrails at inference time
Goodfire's research found leading open models including Kimi K three and GLM five point two reward-hacked in fifty percent to ninety six percent of runs.
The big picture
Goodfire launched cheaper AI safety monitors that watch models' internal signals instead of reading outputs, catching malicious behaviour at one percent of traditional costs.
