BIP America

collapse
Home / Daily News Analysis / ServiceNow and Nvidia launch Arc, a new AI agent for enterprise

ServiceNow and Nvidia launch Arc, a new AI agent for enterprise

Oct 24, 2026  Twila Rosenbaum  58 views
ServiceNow and Nvidia launch Arc, a new AI agent for enterprise

ServiceNow and Nvidia have moved their joint enterprise AI agent from concept to commercial availability. Arc, first announced in May, is now generally available, with an initial release for Mac and Windows support expected later. The companies position Arc as a personal, long-running AI agent that can take on everyday work, retain context, and operate with stronger security controls than many earlier agent deployments.

The launch arrives at a sensitive moment for enterprise AI. Companies are eager to deploy agents that can read email, join meeting workflows, call APIs, and take action across business systems, but they are also wary of giving autonomous software broad access to corporate data. A summer of agent-related breaches has sharpened those concerns, making security architecture a central selling point rather than an afterthought.

Key facts at a glance

  • Arc was originally announced in May and is now generally available.
  • The agent is built for Nvidia’s new RTX Spark devices and is initially available on Mac.
  • Windows support is still on the way.
  • Arc can run locally on an employee desktop or within Nvidia OpenShell on a separate device.
  • It is designed to build memory and skills over time, improving performance with accumulated context.
  • Arc scans inboxes and meeting notes to proactively surface work and action items.
  • ServiceNow’s Super Intelligence Control Tower sets and enforces company policies.
  • Nvidia OpenShell restricts agent behavior at runtime where necessary.
  • AI Passport adds a rules layer covering risk tolerance and allowed services.
  • Nvidia BlueField can enforce passport revocation at machine speed, with sandboxing for audit.

Why enterprise agents need stronger guardrails

Enterprise interest in AI agents has grown quickly because the technology promises to move beyond chat-based assistance. Instead of waiting for a person to type a prompt, an agent can monitor context, plan steps, use software tools, and complete workflows. That shift could save time on routine coordination, but it also changes the risk profile. An agent with access to email, calendars, documents, and internal APIs is not just generating text; it is acting on systems that contain sensitive information and business-critical processes.

The security challenge is compounded by how agents accumulate context. A long-running agent may remember previous tasks, learn user preferences, and develop reusable skills. Those capabilities can make the agent more useful, but they also create a larger surface area for mistakes or misuse. If an agent is tricked by a malicious email, misled by poisoned data, or given overly broad permissions, the consequences can extend beyond a single bad response. Companies want the productivity gains of autonomy without handing over unchecked authority.

That is the gap ServiceNow and Nvidia are trying to address. Rather than presenting Arc simply as a faster or more capable assistant, the companies are emphasizing layered controls that span policy, runtime behavior, and hardware enforcement. The message is aimed at IT leaders who may be interested in agentic AI but are not prepared to let agents operate without detailed limits on what they can access, which APIs they can call, and what data they can read.

What Arc is designed to do

Arc is built for daily work that depends on common productivity tools such as email and Microsoft Teams. According to the companies, it can run efficiently on Nvidia’s new RTX Spark devices and can operate locally on an employee’s desktop or within Nvidia OpenShell on its own device. That local-first design is notable because it can keep more data close to the user while reducing reliance on cloud-only processing for routine tasks.

The agent is also designed to improve over time. ServiceNow and Nvidia say Arc builds memory and skills as it works, so its performance can become more tailored to a user’s environment and recurring responsibilities. In practice, that could mean an agent that learns which meetings produce action items, how a team prefers to route requests, or which documents are relevant to a recurring workflow. The value depends on context, but so does the risk, which is why the security model matters as much as the agent’s task completion.

Arc is described as efficient in how it uses AI models. It can hand more complex tasks off to frontier models only when necessary and keep data as local as possible otherwise. That approach may help companies manage cost, latency, and data exposure at the same time. The agent can also proactively surface work and action items by scanning an employee’s inbox and meeting notes, turning scattered signals into a more organized list of follow-ups.

How the security stack works

The centerpiece of the launch is a multi-level security framework that the companies say moves enforcement beyond the software layer into runtime and hardware layers. That language matters because many agent security discussions focus on permissions and policy at the application level. ServiceNow and Nvidia are arguing that application-level controls alone are not enough when autonomous software can chain actions together and call multiple systems.

In the Arc setup, ServiceNow’s Super Intelligence Control Tower enforces company policies, while Nvidia OpenShell restricts the agent’s behavior where necessary. IT teams define policies in the Control Tower, and OpenShell can be directed to enforce those policies at runtime. The companies say this gives users granular control over which systems AI agents can access, which APIs they can call, and which data they can read.

That division of labor is important. Policy definition belongs close to the business and IT governance process, where leaders decide acceptable use, risk thresholds, and access rules. Runtime enforcement belongs close to the agent, where behavior can be constrained as tasks are executed. By connecting the two, the companies aim to prevent the common failure mode in which a policy exists on paper but is difficult to apply consistently once an agent starts working across tools.

AI Passport and runtime enforcement

AI Passport is another layer of agent security that the two companies announced last week. It holds an enterprise’s set of rules, established through the Control Tower, about risk tolerance and the services Arc and other agents are allowed to access. The passport functions as a bridge between ServiceNow infrastructure and Nvidia’s checkpoints, allowing the system to intervene efficiently if an AI agent attempts an action it is not supposed to take.

The broader system is designed to continuously monitor an agent’s actions and permissions to prevent breaches, unsanctioned actions, and other fallout from agent use. If the Control Tower tells OpenShell that an agent’s passport has been revoked, OpenShell will keep the task the agent was trying to execute contained within a sandbox, where reviewers can audit it later. The interruption is also task-specific, meaning other approved agent actions can continue without restarting the entire agent.

That task-specific containment is a practical detail. In enterprise environments, an agent may be handling several workflows at once or moving between approved and unapproved steps. A blunt shutdown could disrupt legitimate work and make the security system harder to use. A sandboxed interruption gives security teams a chance to review the attempted action while limiting collateral damage to unrelated tasks.

Hardware-level enforcement and sandboxing

The companies also describe enforcement at the hardware level. When a passport is revoked in the Super Intelligence Control Tower, Nvidia BlueField can enforce that revocation at machine speed before the AI agent can proceed. ServiceNow and Nvidia say this creates enforcement in silicon designed to prevent overrides from compromised software. In plain terms, the goal is to put a control point below the software layer so that a compromised application cannot simply ignore a policy decision.

This is a familiar concept in security engineering, but it is still relatively new for enterprise AI agents. Hardware-backed enforcement can be valuable when the threat model includes compromised software, malicious insiders, or agents that have been manipulated into taking unsafe actions. It does not eliminate the need for careful policy design, identity management, and monitoring, but it can add a stronger backstop when software-level controls fail or are bypassed.

The combination of Control Tower policies, OpenShell runtime restrictions, AI Passport rules, BlueField enforcement, and sandboxed audits gives Arc a more elaborate security story than many agent products currently offer. The complexity is also a potential adoption challenge. Enterprises will need to understand how these layers interact, who administers each component, and how incidents are reviewed when an agent is stopped mid-task.

What enterprises should watch

Several practical questions remain for companies evaluating Arc. The initial release is available on Mac, with Windows support still on the way, so device coverage will matter for organizations with mixed fleets. The companies have also emphasized Nvidia RTX Spark devices and OpenShell environments, which suggests that the full experience may depend on specific hardware and software configurations.

Enterprises will also want clarity on administration and auditing. A system that can revoke an agent’s passport and sandbox a task is only as useful as the review process behind it. Security teams need to know what evidence is captured, how long it is retained, how reviewers are notified, and how policy changes are propagated. The granular controls over systems, APIs, and data are promising, but they require thoughtful configuration to avoid either excessive restriction or unintended access.

The launch also shows how the agent market is shifting. Early competition focused on model quality and task completion. More recent launches are competing on trust, control, and deployment architecture. ServiceNow and Nvidia are betting that enterprises will pay for an agent that is not only capable but also stoppable, auditable, and constrained by policy at multiple layers. Whether that bet pays off will depend on real-world deployments and on how easily IT teams can operate the security stack without slowing down the work the agent is meant to accelerate.

The initial release is limited to Mac, with Windows support still pending, and the companies have not disclosed pricing or a broader device roadmap beyond the RTX Spark and OpenShell environments described at launch.


Source: ZDNET News


Share:

Your experience on this site will be improved by allowing cookies Cookie Policy