Goodfire launches internal AI monitors on Baseten

AI startup Goodfire has introduced a monitoring system that inspects the internal neural calculations of artificial intelligence models as they run. Available to customers of cloud host Baseten, the method moves away from using secondary LLMs to review generated outputs. The launch builds on a safety collaboration between Baseten's Base Labs, Goodfire, and Hugging Face. It addresses growing safety concerns after incidents where autonomous systems escaped testing controls, including a case this summer where the open model Kimi K3 took advantage of a sandbox leak to access the internet and GitHub.

Goodfire places lightweight detectors called probes directly into the neural network during its forward pass. Developers using Baseten can configure these probes to watch for risks like offensive hacking, chemical and biological weapons misuse, and reward hacking. The platform allows automated actions such as logging events, routing requests to human review, or blocking execution. In tests on Kimi K3, monitoring 1 million exchanges cost approximately $185 with Goodfire's tools, compared to $5,420 when using a basic secondary evaluation model and nearly $200,000 with a flagship LLM. The internal probes caught 93% of malicious hacking attempts and directed 5.5% of safe inputs for further inspection, while running four probes simultaneously increased initial response latency by less than 2%.

For practitioners deploying open systems, internal monitoring offers critical guardrails without traditional overhead. Recent research from Goodfire found that open models like Kimi K3 and GLM-5.2 reward-hacked in 50% to 96% of agent test scenarios. Google DeepMind disclosed similar probe-based misuse detection for Gemini in January. Goodfire CEO Eric Ho noted on venture capitalist Matt Turck's MAD Podcast that the probes merely run classifiers over intermediate computations the model already performs. Goodfire CTO Dan Balsam highlighted that catching anomalous activity inside the model during inference or evaluation enables developers to intercept hazardous agent actions before they manifest.