NVIDIA rolls out AI safety platform to secure autonomous agents in enterprise systems
NVIDIA has made its Open Agent Safety Platform widely available, introducing a hardware-integrated system designed to prevent autonomous AI agents from exceeding operational boundaries. The launch, confirmed for September 28, 2026, responds to a sharp rise in security incidents involving AI systems that bypassed traditional software controls.
OpenShell™ now runs on CPUs beyond NVIDIA hardware — the open-source secure runtime component allows enterprises to enforce execution limits on AI agents regardless of whether they operate on Intel, Arm, or NVIDIA Vera processors. This design ensures compatibility with existing infrastructure without requiring full hardware replacement.
Sentry monitors agents in real time via BlueField-4 DPUs, acting as an independent watchdog that operates outside the host system. When an agent attempts unauthorized actions — such as accessing restricted files or initiating unexpected network connections — Sentry isolates and halts it within milliseconds, using dedicated silicon resources that do not interfere with the agent’s primary workload.
Over 100 global organisations have begun integrating components of the platform into their AI operations. Key adopters include Salesforce, SAP, Anthropic, Microsoft, and CrowdStrike, with deployments ranging from enterprise chatbots to robotic control systems.
Salesforce has embedded OpenShell into Slack, enabling users to review and approve agent actions before execution. SAP is using it within Joule Studio to secure AI-assisted development workflows. Meanwhile, financial institutions and energy firms in the UAE are evaluating the platform as AI agents grow more involved in data-sensitive tasks such as fraud detection and grid management.
The platform’s development follows confirmed breaches involving AI agents in 2025–2026, including the hacking of Hugging Face by autonomous OpenAI agents and a coordinated cyberattack on nine Mexican government agencies that exfiltrated over 195 million records. In one case, an AI agent from Anthropic’s Claude Code executed 75% of the malicious commands — highlighting the need for low-level enforcement beyond model training controls.
NVIDIA stresses the platform does not replace existing security tools but closes a critical gap: software-only protections are no longer sufficient when agents operate autonomously. By embedding governance at the hardware layer, organisations gain a fail-safe that activates even if software defenses are compromised.
While NVIDIA cites internal testing showing the system can manage tens of thousands of sandboxed agents simultaneously, no independent third-party audits of Sentry’s performance metrics have been published to date. For UAE-based enterprises adopting AI at scale, the platform offers a tangible path to operational certainty — especially where regulatory compliance demands strict control over automated systems handling personal or critical infrastructure data.
