Base Labs is developing open safety research and monitoring methods to be integrated into Baseten's inference infrastructure. The initiative, created with collaborators Hugging Face and Goodfire AI, provides training for models to follow explicit policies and includes runtime failure detection with linked intervention controls to create a transparent standard for how open AI models are deployed.

Baseten will offer the safety infrastructure as a managed service to its customers to establish security standards for open-source models. The effort follows an agent swarm attack by OpenAI on Hugging Face, which Baseten describes as the type of failure providers must prepare for as open models reach parity with closed frontier labs.