Goodfire Launches Internal AI Agent Monitors on Baseten

Goodfire AI launched internal activation monitors for AI agents through Baseten, according to TechCrunch. The monitors inspect signals inside a model rather than using a separate model to review every output.

Customers can monitor risks including offensive hacking, chemical and biological weapons misuse, and reward hacking. They can configure the system to log flagged activity, send it for human review, or refuse a request.

In tests using Kimi K3, Goodfire said monitoring about 1 million exchanges cost roughly $185. The probes detected 93 percent of malicious hacking sessions and flagged 5.5 percent of harmless sessions for further review. Running four probes added less than 2 percent to the time before the model began responding.