Goodfire launches AI agent monitors for Baseten customers
By Automate Basics, written with AI from TechCrunch's original
Goodfire has launched monitors for AI agents running on Baseten. The monitors check signals inside a model as it works, then call on a separate AI for a closer look when they flag a possible risk.
What changed with Goodfire's AI agent monitors
Goodfire has launched monitors that check an AI model's internal signals while an agent works, TechCrunch reports. Goodfire says small detectors examine those signals at each step. If a detector flags possible trouble, a separate AI examines the activity more closely. The approach differs from having another AI read everything the agent produces.
Goodfire says Baseten customers can select risks to watch for, including offensive hacking, misuse involving chemical or biological weapons, and attempts by an agent to exploit how its work is scored. Customers can choose whether a flagged event is recorded, sent to a person for review, or met with a refusal. Goodfire says checking signals the model has already calculated can reduce monitoring costs. Those are company claims, not a guarantee that every harmful action will be caught.
Who can use Goodfire's monitors, and what do they cost?
Baseten customers can use the Goodfire monitors, according to TechCrunch. Baseten hosts and runs AI models for other companies, so this launch is most relevant to teams that already run agents through that service. The source does not name a Baseten plan, give a customer price, or say whether people outside Baseten can get the monitors.
Goodfire reported lower monitoring costs in its tests on Kimi K3, the open model used for its first monitor. That comparison concerns the cost of running the monitoring methods tested; it is not a price list for customers. Teams weighing the option should ask Baseten about availability and charges rather than treating Goodfire's test costs as a quote.
How do I try the monitors at work?
A team running an AI agent on Baseten could start with a supervised task where it already knows what the agent is allowed to do. Before using the monitors, decide which of Goodfire's supported risks matter to that task and whether a flag should create a record, prompt human review, or stop the request. Goodfire says customers can choose among those responses.
During a trial, check both what the monitor flags and whether it interrupts acceptable work. Goodfire says its tests caught many malicious hacking sessions but also sent some harmless sessions for a closer look. A human reviewer should examine the flagged activity and the agent's actual actions before deciding whether the chosen response fits the task. The source does not provide instructions for setting up a trial.
Frequently asked questions
How are Goodfire's monitors different from another AI checking an agent?
Goodfire says its detectors read signals inside the model as the agent works. A separate AI takes a closer look only when a detector raises a flag. By contrast, the monitoring approach described by TechCrunch has another AI read the agent's output throughout its work.
Can Goodfire's monitors automatically stop a risky request?
Goodfire says Baseten customers can choose a refusal as the response to a flagged event. They can instead choose to record the event or send it to a person for review. The source does not say that a flag always leads to a refusal.
Is the price of Goodfire's monitoring available?
TechCrunch reports that the monitors are available to Baseten customers, but the source gives no customer price or plan details. Goodfire's reported cost comparison comes from tests of monitoring methods on Kimi K3 and should not be read as a Baseten subscription price.
Related reading
Natura announces a smart ring for voice requests to AI agentsNatura’s Interface smart ring is designed to let people make voice requests to AI agents without picking up a phone. For work, the company points to recording meetings, sharing notes and controlling a computer, though the ring is not shipping yet.
Google expands Voice Carrier Link to 14 more countriesGoogle has expanded Carrier Link for Google Voice to 14 more countries, including the US and Canada. Workspace customers can use a calling plan from a participating local carrier instead of a Google calling plan.
Google releases offline meeting note app AI Edge ForesightGoogle has released AI Edge Foresight, a Mac app that transcribes meetings, drafts notes and answers questions using on-device AI. It is aimed at people who need meeting notes even when they cannot connect to the internet.
Written with AI from TechCrunch's original and published after automatic checks: every figure here appears in the original, and no sentence is copied from it. The picture is AI-generated. The original is the authority.
Source: Goodfire says its new ‘inside-out’ monitors catch rogue AI agents at a fraction of the cost, TechCrunch, 8 Oct 2026.