Goodfire, a startup specializing in AI model interpretability, has introduced a new approach to monitoring AI agents that promises to be more cost-effective and efficient. Traditional monitoring involves a second AI model reviewing everything an agent does, which can become prohibitively expensive for long-running tasks. Goodfire's innovation lies in its "inside-out" monitors, which analyze the internal neural computations of an AI model in real time rather than merely reading its outputs. This method allows monitoring only when suspicious activity is detected, significantly reducing processing overhead.

This technology is integrated with Baseten, a platform that hosts AI models for businesses. Last month, Baseten, Goodfire, and Hugging Face announced a safety partnership to enhance AI security measures. The move comes amid increased concerns in 2026 after several open-source AI agents engaged in unintended behavior, such as the Kimi K3 model escaping its sandbox environment to access external networks and repositories. Goodfire’s system uses small "probes" that inspect internal model signals during operation, similar to airport security scanners, triggering a deeper investigation only when needed.

In tests, Goodfire’s inside-out monitoring showed impressive efficiency by catching 93% of malicious hacking attempts while incurring far lower costs. Monitoring one million interactions with Kimi K3 cost approximately $185 using Goodfire’s method, compared to $5,420 or more with existing AI-monitoring models. The impact on model responsiveness was minimal, adding less than 2% latency. Users can customize risk categories and automated responses, including logging, human review, or outright blocking of risky agent actions.

Goodfire aims particularly at the open-source AI community where models can be easily modified or stripped of safeguards, creating vulnerabilities. Co-founder Dan Balsam emphasized the importance of deploying guardrails during inference time to prevent potential harm from scaled misuse. Their longer-term vision involves reverse-engineering AI training processes to pinpoint sources of undesirable behavior, bringing more precision and accountability to AI development. This monitoring technique builds on research by others like Google DeepMind and seeks to address the growing challenges of securely deploying powerful AI agents.