Jensen Huang on AI safety: contain agents, monitor them in real time with BlueField chip
XFreeze · x · 2026-09-30
Jensen Huang laid out NVIDIA's approach to keeping AI safe: models are relatively safe during training, but evaluation and testing require a highly secure containment system for the agentic stack — and you should never assume containment holds, so monitoring must run in real time on separate hardware outside the agent's sandbox. The same containment-plus-monitoring model applies after deployment, analogous to onboarding a new employee with limited rights and continuous oversight. NVIDIA introduced two technologies for this: OpenShell, a containment system described as "a browser for agents" that isolates the agent from the rest of the computer with minimal permissions, and BlueField, a chip that watches everything the agent does and escalates on out-of-bounds behavior. Huang argued safety and advancing the technology are "one and the same thing," dismissing capability-vs-safety tradeoffs as a false choice.
More from Infra
- AT&T CEO says SpaceX's phone strategy is not viable: "Satellite won't beat fiber" — RachelVT42 · 2026-09-30
- ModelScope CLI quietly moved to the modelscope-hub package — PSA to save you 30 minutes — pilkyton · 2026-09-30
- Rowmax-H15: approximate softmax in attention for 25.8% faster B200 inference with minimal quality loss — illinois · 2026-09-30
- AWS Bedrock Brings Claude Opus 5, Sonnet 5, Haiku 4.5 to In-Country Inference in India — AWS ML Blog · 2026-09-30
- Bedrock Adds In-Region Claude Inference in Seoul and Singapore for Data Residency — AWS ML Blog · 2026-09-30
- All PyTorch Conference China 2026 sessions now available on YouTube — PyTorch · 2026-09-30