BIP America News & Media Platform

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

ServiceNow and Nvidia have made Arc generally available on Wednesday, turning a concept first announced in May into a commercial product for companies that want AI agents to handle everyday work without giving up control over data, permissions, and execution. Arc is described by the two companies as a personal long-running AI agent that can take on routine enterprise tasks, build memory and skills over time, and operate with security controls that extend beyond the software layer into runtime and hardware enforcement.

The launch lands at a tense moment for enterprise AI. Companies are eager to deploy agents that can read email, join meetings, summarize notes, and take action across productivity tools, but they are also wary of autonomous systems that can access sensitive systems, call APIs, and move data without adequate oversight. A summer of agent-led breaches has sharpened those concerns, making security architecture a central selling point rather than an afterthought.

Key facts at a glance

  • Arc is now generally available after being announced in May.
  • The agent is built for Nvidia's new RTX Spark devices.
  • Arc is available on Mac at launch, with Windows support still to come.
  • It is designed as a personal long-running AI agent for everyday enterprise work.
  • Arc can run locally on an employee desktop or inside Nvidia OpenShell on its own device.
  • The agent builds memory and skills over time to improve performance.
  • ServiceNow and Nvidia emphasize multi-level security, with enforcement moving beyond software into runtime and hardware layers.
  • ServiceNow's Super Intelligence Control Tower sets and enforces company policies, while Nvidia OpenShell restricts agent behavior when needed.
  • AI Passport, announced last week, holds enterprise rules on risk tolerance and allowed services.
  • If a passport is revoked, OpenShell can contain a task in a sandbox for later audit while other approved actions continue.
  • Nvidia BlueField can enforce revocation at machine speed in silicon.

Why security is the real product story

For all the attention that new agent capabilities attract, the harder problem for enterprise adoption is control. An AI agent that can draft messages, schedule meetings, search internal documents, and interact with business applications is useful precisely because it has reach. That reach also creates risk. An agent with excessive permissions can become a pathway for data leakage, unauthorized transactions, or prompt-injection attacks that trick it into taking actions outside its intended scope. Security teams want visibility into what an agent can access, which APIs it can call, what data it can read, and how quickly a bad action can be stopped.

Arc's launch is notable because ServiceNow and Nvidia are not positioning security as a dashboard bolted onto an agent after the fact. They describe a layered model in which policy, runtime enforcement, identity-like credentials, and hardware-level controls work together. The companies said their approach moves enforcement beyond the software layer into runtime and hardware layers, a distinction that matters when software itself may be compromised or when an agent tries to override its boundaries.

The reference to agent-led breaches is important context. Enterprise interest in autonomous agents has grown quickly, but so have concerns about the gap between demonstration and dependable deployment. Agents often need broad access to email, calendars, chat, documents, and internal systems to be useful. Without fine-grained controls, that access can turn a productivity assistant into a lateral-movement risk. The security question is not only whether an agent makes a mistake, but whether the surrounding infrastructure can prevent, contain, and audit that mistake before it spreads.

What Arc can do

Arc is designed for daily tasks that rely on productivity tools such as email and Microsoft Teams. It is built to run efficiently on Nvidia's new RTX Spark devices, which means it can execute locally on an employee's desktop or within Nvidia OpenShell on its own device. That flexibility is significant. Local execution can reduce the amount of sensitive data sent to external models, while a dedicated agent device or sandbox can isolate work from an employee's primary machine.

Like other enterprise agents that use accumulated context, Arc is intended to build memory and skills over time. Instead of treating every request as a fresh interaction, it can develop a working understanding of an employee's tasks, preferences, and recurring workflows. The companies said this allows performance to improve continually. Arc also proactively surfaces work and action items by scanning an employee's inbox and meeting notes, a capability that positions it closer to an assistant that monitors workflow than a chatbot that waits for prompts.

The agent is also designed for efficiency. ServiceNow and Nvidia said Arc hands more complex tasks to frontier models only when necessary and keeps data as local as possible otherwise. That routing strategy matters for cost, latency, and privacy. Many enterprise tasks are repetitive and context-dependent but not intellectually complex. Running them locally or on smaller models can be cheaper and faster than sending every request to a large cloud model. When a task requires stronger reasoning, the agent can escalate to a frontier model, but the default is to minimize data movement.

Arc's availability on Mac at launch, with Windows support still on the way, suggests an initial focus on knowledge workers and developer-friendly environments. Macs are common among technical and creative employees, and local AI workloads can benefit from modern hardware acceleration. Windows support will be important for broader enterprise deployment, given the size of the Windows installed base in large organizations.

How the security stack works

The centerpiece of the ServiceNow and Nvidia partnership is a multi-level security framework around Arc. ServiceNow's Super Intelligence Control Tower is used to define and enforce company policies. Nvidia OpenShell restricts the agent's behavior where necessary. In practice, IT teams can set rules in the Control Tower, and OpenShell can be directed to enforce them at runtime. The companies said this gives users granular control over which systems AI agents can access, which APIs they can call, and what data they can read.

That runtime enforcement is a meaningful departure from static permission models. Traditional access control often grants or denies access at the point of login or API call, but an agent may chain together many actions across different systems. Runtime controls can evaluate what the agent is trying to do while it is doing it, rather than assuming that initial permissions are sufficient. For enterprises, that could help contain risky behavior such as an agent attempting to export a large set of documents, contact an unauthorized external service, or escalate its own privileges.

AI Passport adds another layer. Announced last week, it holds an enterprise's set of rules about risk tolerance and the services Arc and other agents are allowed to access. Those rules are set by the Control Tower. AI Passport functions as a bridge between ServiceNow infrastructure and Nvidia's checkpoints. The system is designed to monitor an agent's actions and permissions continuously to prevent breaches, unsanctioned actions, and other fallout from agent use.

The revocation workflow is especially notable. If the Control Tower tells OpenShell that an agent's passport has been revoked, OpenShell keeps the task the agent was trying to execute contained within a sandbox, where reviewers can audit it later. The interruption is task-specific, meaning other approved agent actions can continue without restarting the entire agent. That granularity could reduce operational disruption. Instead of shutting down an agent completely after a suspicious action, security teams can isolate the specific task, preserve evidence, and allow legitimate work to proceed.

Nvidia BlueField adds hardware-level enforcement. The companies said that 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. They described this as enforcement in silicon designed to prevent overrides from compromised software. Hardware enforcement is particularly relevant for agents that run on dedicated devices or edge infrastructure, where software-only controls may be bypassed if the operating system or agent runtime is compromised.

Local execution, frontier models, and enterprise trade-offs

Arc's architecture reflects a broader shift in enterprise AI toward hybrid execution. Sensitive context can stay local, while heavier reasoning can be routed to frontier models when needed. This is not just a privacy feature. It is also a cost and latency strategy. Large models are powerful but expensive and slower for routine work. A long-running agent that scans inboxes, tracks action items, and updates workflows needs to operate continuously without generating unsustainable inference bills. Local processing and selective escalation offer a practical middle ground.

At the same time, local execution introduces management challenges. IT teams must provision devices, manage updates, monitor agent behavior, and ensure that local data stores are protected. Running an agent within Nvidia OpenShell on its own device could simplify isolation, but it also creates another endpoint to secure. The combination of Control Tower policies, AI Passport rules, OpenShell runtime restrictions, and BlueField enforcement is intended to address those concerns, but enterprises will need to validate how the pieces work in their own environments.

The partnership also highlights how infrastructure vendors are positioning themselves for the agentic era. Nvidia is supplying compute, runtime isolation, and hardware enforcement, while ServiceNow is supplying policy control and enterprise workflow context. That division of labor could appeal to companies that want agent capabilities without building every security layer from scratch. It also raises integration questions: how well will Arc connect to existing identity providers, data loss prevention systems, endpoint management tools, and security information and event management platforms? The answers will determine whether Arc becomes a manageable part of enterprise operations or another silo to monitor.

What enterprises will watch next

Early adopters will likely focus on a few practical questions. Can Arc reliably surface action items from email and meeting notes without creating noise? How much memory does it accumulate, and how is that memory governed? Which tasks can run locally, and when does the agent escalate to a frontier model? How quickly can a revoked passport stop a risky action, and what evidence is preserved for auditors? Windows support will also be a key milestone for broad deployment.

The security architecture will face real-world testing. Agent behavior is probabilistic, and attackers are already probing ways to manipulate agents through prompt injection, malicious documents, and compromised integrations. A multi-level framework that combines policy, runtime restrictions, sandboxing, and hardware enforcement gives defenders more layers to work with, but no system eliminates risk entirely. The value of Arc will depend on whether those layers are usable, transparent, and fast enough for everyday work.

For now, the launch marks a notable step in the commercialization of long-running enterprise agents. The pitch is not simply that Arc can do more, but that it can do more within guardrails that reach from policy definitions to silicon. As companies move from agent experiments to production deployments, that combination of capability and control is likely to matter more than any single feature. Windows support, integration with existing security tooling, and real-world testing of the revocation workflow will be the next practical milestones for enterprise buyers.


Source: ZDNET News


Share:

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