NVIDIA Launches Open Agent Safety Platform
NVIDIA has announced the release of its Open Agent Safety Platform, a new architectural framework designed to secure autonomous artificial intelligence agents at the infrastructure level. According to MarkTechPost, the platform introduces OpenShell, a tool that sandboxes software agents on Vera CPUs, restricting unauthorized system access during execution.
In addition to OpenShell, the platform incorporates Sentry, a security mechanism running on BlueField-4 data processing units. Sentry is engineered to monitor agent network activity and isolate compromised instances within milliseconds if anomalous behavior or security violations are detected.
As builders increasingly deploy autonomous multi-agent workflows capable of executing system commands and interacting with external APIs, hardware-level isolation has emerged as a critical requirement. The integration of CPU sandboxing and network-level quarantine aims to prevent runaway loops, prompt injection exploits, and lateral movement across enterprise networks without introducing prohibitive runtime latency.
The initiative targets developers and infrastructure architects building production-grade agentic systems who require robust containment strategies. Additional details regarding software development kits, implementation guidelines, and availability timelines are expected to be released through NVIDIA developer channels.
Based on reporting by www.marktechpost.com.
