Nvidia launches Open Agent Safety Platform to restrain rogue AI agents — new hardware-and-software security stack can quarantine agents in milliseconds - Tom's Hardware

Nvidia's decision to embed agent quarantine at the GPU microarchitecture level is a strategic repositioning from compute vendor to trust-infrastructure provider for the agentic AI era. Millisecond hardware interrupts are fundamentally harder to bypass than any software sandbox — this is a behavioral constitution etched into silicon. The technical ripple is immediate: inference frameworks must expose safety API hooks, FPGA co-processors require redesign, and the entire AI middleware stack faces structural restructuring. On compliance, the EU AI Act's auditability mandates and emerging US state-level agent liability laws converge on one hard requirement: hardware-level, tamper-proof behavioral logs. Without this layer, your AI agents are black boxes in regulatory audits, and enterprise insurance premiums spike accordingly. Competitively, AMD's CDNA architecture lacks an equivalent safety interrupt mechanism near-term. The real threat, however, comes from hyperscalers — AWS and Azure will likely ship proprietary agent isolation layers within 12 months to circumvent what amounts to Nvidia's "safety tax." Open-source RISC-V alternatives remain at least 18 months behind in ecosystem maturity. Over the next 12–24 months, agent safety will shift from optional to mandatory — analogous to how Secure Boot reshaped the PC industry. Leveraging CUDA's lock-in, Nvidia will almost certainly cement this layer as the de facto standard, merging compute pricing power with safety pricing power into a single moat.