The standard way to keep an AI agent in check is to hire a second AI to read everything the first one does. It works, but it's like paying a security guard to re-read every email your staff sends. Goodfire has a cheaper plan: look inside the agent's head instead.
Probes Inside the Model, Not After It
On October 8, 2026, San Francisco interpretability startup Goodfire launched monitors that read an AI model's internal activations while it works, rather than reviewing its output afterward. These "activation probes" flag behaviors such as offensive hacking, chemical or biological weapons misuse, reward hacking, and suspicious tool use. When something looks off, the system can log an alert, send the session to a human reviewer, or refuse the request.
The numbers are striking. In tests on Kimi K3, monitoring 1,500 sessions cost about $51, compared with roughly $233 for cheaper model-based review and around $10,000 for premium review. The probes flagged 94% of malicious hacking sessions while escalating 8.7% of harmless ones, and running four probes at once added less than 2% to response time. The monitors are live now for customers on Baseten's inference platform.
Security That Scales With Agents
This matters because agents are multiplying faster than anyone's review budget. If watching an agent costs more than the agent itself, teams will quietly skip the watching, and that's how things go wrong. Pennies per session makes monitoring a default setting instead of a luxury.
A reality check is in order, though. Probes need access to model internals, so for now this works where you control or host the model, and availability beyond Baseten is unclear. Agents you rent through a closed API are still a black box. Still, an 8.7% false-positive rate that triggers a second look, not a shutdown, is a sensible design.
The bigger shift is philosophical. Most AI safety tooling judges what a model says. Goodfire is betting you can catch trouble while the model is still deciding what to do, before a bad tool call reaches your database or inbox. Separate tests on GLM-4.5-Air showed 97% recall at a 1% false-positive rate, which improved to 99% after more iterations. If those results hold up across more models, interpretability research stops being an academic curiosity and starts being a security product with a price tag.
Trust, but verify, preferably for three cents a session.
If your business is adding AI agents to its website or workflows and you want guardrails built in from the start, let's map out a safer setup together.
Source: Runtime Wire