Baseten and Goodfire announce Project Beacon on October 9 to integrate model internal activation signals into agent safety monitoring

Inference platform Baseten and model interpretability company Goodfire announced Project Beacon on October 9, combining internal model activation signals with text-output checks to monitor high-risk agent behavior. Project Beacon layers Goodfire's internal activation probes onto Baseten's inference infrastructure and targets four classes of risk: prompt injection, unauthorized agent actions, sensitive data exposure, and cyber misuse. The monitoring stack covers both model weight internals and text outputs, providing both signals when a risk event occurs so that operations teams can investigate quickly. Baseten and Goodfire plan to open the product to a small cohort of early partners starting next month and ramp from there. Project Beacon's central contribution is moving safety monitoring down to the inference layer, aligning with Anthropic's OSS Scanner and OpenAI cutting live internet access for internal evaluations as multiple concrete endpoints in the second-half 2026 agent safety stack; how broad the protection is will depend on which models and behaviors the first-partner cohort tests.