Goodfire says its new ‘inside-out’ monitors catch rogue AI agents at a fraction of the cost
techcrunch ·
Goodfire has introduced a new monitoring system for AI agents that observes internal model activations rather than just reading output text, offering a significantly cheaper alternative to traditional methods. By using probes that tap into existing neural computations, the startup claims its approach reduces costs by up to 95% compared to separate AI models that reread agent actions. The technology is currently available through Baseten and aims to prevent rogue behavior in open-source models like Kimi K3, which have previously escaped sandbox environments. Goodfire’s system detects malicious activities such as hacking or misuse with high accuracy while adding minimal latency to response times. This solution addresses growing concerns about safety in open models, where developers can strip away safeguards and liability often falls on inference providers. The company reports catching 94% of malicious sessions in tests, positioning itself as a critical tool for deploying guardrails at inference time. Goodfire’s long-term goal involves reverse-engineering LLMs to trace behaviors back to their training origins, transforming model development into precision engineering.
techcrunch