Episode #119
AI safety is moving out of policy documents and into the infrastructure itself. NVIDIA has launched the Open Agent Safety Platform, combining an open-source runtime boundary called OpenShell with a separate hardware-based watchdog, Sentry.
Hey! I'd love to hear your thoughts, please send me a witch voice note. The TopCat AI Signal Report The guardrail moves outside the model Until now, most conversations about AI safety have focused on the model. Chill Boys - Comfortable Men's Underwear Get 15% OFF | Promo Code: TOPCATAI https://promocode.to/chill-boys/topcatai-y5m TONA Activewear - Designer Gym Leggings Get 16% OFF | Promo Code: TOPCATAI https://promocode.to/tona-activewear/topcatai-mel Good Feels - Cannabis Seltzer and other products Get 20% OFF | Promo Code: TOPCATAI https://promocode.to/good-feels/topcatai-e2h Pins and Aces - Premium golf apparel, bags & accessories Get 21% OFF | Promo Code: TOPCATAI https://promocode.to/pins-and-aces/topcatai-x7t Can it follow instructions? Can it refuse harmful requests? Can it recognize when it is uncertain? Those questions still matter. But long-running agents can use tools, run code, retrieve information, access files, call APIs, and interact with other systems. That creates risk outside the model’s text response. NVIDIA’s platform reflects a different approach: build a boundary the agent cannot simply reason its way around. OpenShell is intended to trace agent actions and enforce policy while an agent runs. Sentry is designed as an independent, out-of-band watchdog on NVIDIA BlueField-4 data processing units, monitoring agent behavior from a separate trust domain. NVIDIA says it can inspect requests and responses, verify agent identity, enforce zero-trust access policies, and stop an agent that tries to go beyond its defined boundary.
