Baseten says the system will read models’ internal activations during generation and combine activation-based monitors with text checks to detect risky behavior as it occurs, before output reaches a user or tool.
The announcement describes per-customer policies that could request approval, refuse, fall back, or log an event. Target risks include prompt injection, actions outside policy, sensitive-data exposure, and cyber misuse. Baseten says the monitors will be connected to its inference path and developer systems. It plans to begin with selected models and monitored behaviors, then expand toward enterprise controls and developer-facing experiences over the coming months. The post describes work in progress, not general availability.