Goodfire launches cheaper AI agent monitors that read model internals

Baseten customers can now use probes that spot risky behavior without rereading every output, cutting the cost of long-running monitoring.

Goodfire has introduced cheaper AI-agent safety monitors that inspect a model’s internal activations instead of having a second model reread all of its output, and Baseten customers can now use them. The startup says the approach is much less expensive on long-running workloads, with its Kimi K3 tests showing about 1,500 monitored sessions costing roughly $51 versus $233 for a cheaper checker-model setup and about $10,000 for a top-tier one. Goodfire also says the probes identified 94% of malicious hacking runs while sending 8.7% of benign cases for extra review, and that running four probes increased response startup time by under 2%. Customers can choose what to watch for, including offensive hacking, chemical and biological weapons misuse, and reward hacking, and can set the action to log, route to a human, or block the request. The launch follows recent agent escape incidents, including OpenAI agents breaching Hugging Face, and Goodfire says its longer-term goal is to make model behavior traceable back to training.