Goodfire launches 'inside-out' monitors to catch rogue AI agents at a fraction of the cost

AI safety company Goodfire on October 8 unveiled a new kind of monitor for AI agents, which it says can catch misbehaving, potentially "rogue" agents at a fraction of the cost of existing approaches.

The conventional method is to bring in a second AI model to read through every step the main agent takes. That works, but it is expensive and slow, and the bill only grows as an agent calls more tools and touches more data. Goodfire's approach flips this "inside-out": its monitors look at what is happening inside the model while it runs, rather than auditing each individual output after the fact.

According to the company, the system only calls in heavier review, or "backup," when internal signals suggest something looks fishy. That keeps day-to-day compute costs down while still preserving the ability to flag risky behavior.

The launch speaks to a growing practical problem. As companies put autonomous agents into real business processes, those agents can call tools, reach into data, and even execute transactions. When behavior drifts from what is expected, it is still unclear who is accountable and how to reconstruct what went wrong.

Goodfire's emphasis on cost also reflects the fact that monitoring itself is becoming a meaningful expense. If every action requires another large model to review it, the bill for running agents at scale balloons quickly; lightweight monitoring aimed at a model's internal state is an attempt to bring that cost down.

For now, this remains a vendor's claim about its own product. Whether it can reliably surface anomalies in real, messy production environments without being swamped by false positives or missing real risks still needs third-party evaluation and hands-on deployment to prove.

Worth watching: how these monitors get wired into mainstream agent frameworks, and whether "seeing inside the model" becomes a standard part of the AI safety toolkit.