Baseten builds safety infrastructure for open models
Baseten and its research arm Base Labs plan to build safety evaluation and monitoring into the infrastructure that trains and serves open models, and to offer it as a managed service.
Inference provider Baseten announced an effort to build safety evaluation and monitoring infrastructure for the open models it serves. Base Labs will lead the research side. The post's subtitle names Goodfire and Hugging Face as collaborators but doesn't say what either will contribute.
The company says it will tie safety research directly to the systems it uses to train and serve models. The goal is to test how training changes affect model behavior, whether monitors detect failures, and whether deployment controls catch and contain them. The plan has three parts:
- Training: teach models to follow explicit policies, recognize their limits, and ask for help when a task can't be completed legitimately.
- Monitoring and intervention: study inputs, outputs, tool use, and internal model states, then link useful signals to permissions, isolation, and checks before execution.
- Learning from failures: investigate incidents, update training or controls, and check that fixes don't get in the way of legitimate tasks.
Base Labs will develop and publish the training and monitoring methods. Baseten will build them into its runtime deployment stack and sell them as a managed service. It says customers and outside evaluators will be able to test the published methods and evidence for themselves.
Baseten argues that people often confuse "open" with "unmonitored." It compares open models to Linux, where openness helps the cycle of finding and fixing vulnerabilities. The company invited other open-model providers to contribute to shared safety evaluation protocols.
The announcement does not give a timeline, pricing, or a list of supported models.