ServiceNow Will Issue Passports to AI Agents, and NVIDIA Will Enforce Them

A credential that travels with the agent

On October 15, 2026, ServiceNow announced AI Passport, a cryptographically signed credential that states exactly what an AI agent is authorized to do inside an enterprise. The passport is checked at runtime at every step of a multi-agent workflow, including inside environments such as NVIDIA OpenShell. It is issued by the company's Super Intelligence Control Tower, the governance product it had previewed to press and analysts as AI Control Tower two days earlier.

Each credential is grounded in the enterprise's live business context: ServiceNow's configuration management database (CMDB), its governance, risk and compliance (GRC) policies, and its workflow approvals. For third-party agents, that creates a verifiable record of permissions instead of relying on an agent's own claims about what it is allowed to touch. The passport layer is built on open specifications, so any governance product, identity provider, or enforcement point can issue and verify credentials for agents built on any vendor's stack.

The trust gap is now a measured number

According to G2's 2026 AI Agents Insights Report, released two days earlier, agent adoption among B2B software buyers rose 25 percent year over year, from 69 percent to 86 percent, and 66 percent of companies now run agents in limited or full production. Four in five organizations using agents in high-volume workflows report cycle-time improvements of at least 11 percent, and more than eight in ten buyers say agents improve employee satisfaction by reducing repetitive work. Only 14 percent report reducing headcount.

But oversight is eating a large share of those gains. Nearly a third of buyers say employees spend at least six hours each week supervising agent work, and 45 percent report losing more than a quarter of expected productivity gains to oversight and rework. G2 calls this overhead the "Distrust Tax," and it hits the largest enterprises hardest.

A separate study from OutSystems, conducted with CIO Dive and KPMG and released October 16, found the same disconnect at the executive level. Eighty-two percent of executives believe their organizations can keep pace with governing AI agents as they move into production. Yet only 26 percent say they are well prepared to resolve conflicts across agents, systems, and automated workflows. In financial services the gap is starkest: ServiceNow's 2026 Banking AI Maturity Index found that while 100 percent of banks report using AI, only 24 percent have governance to match it.

You cannot control what you cannot identify, and you cannot govern what you cannot see. Real governance is graduated: more like a dimmer switch than a kill switch.

Graduated enforcement, not just a big red button

ServiceNow demonstrated the passport with a deliberately mundane example. An out-of-the-box "Order Issues and Refunds" agent from a third-party vendor declared five skills. Control Tower narrowed it to a single job: refund processing for a hypothetical beverage company. Pre-flight checks flagged that the agent had no ceiling on the money it could refund, disabled the four unused skills, and set a $25 limit. Then came two requests. An $18 refund executed automatically. A $40 request was rejected at runtime and routed for human review.

When an agent drifts outside its passport scope, the system can redirect it, restrict it, reconnect it with proper context, or, if termination is required, shut it down at machine speed. A senior ServiceNow AI product executive told the briefing audience that the single red shutdown button makes for good stories and poor oversight: real governance moves through stages. You monitor, you pause, you restrict, and only eventually do you deactivate. Without a signed baseline of what an agent is authorized to do, every guardrail is advisory, because the system has no ground truth to check behavior against.

NVIDIA carries the policy into the runtime

NVIDIA is the design and launch partner for AI Passport, building a production-ready enforcement implementation on its Open Agent Safety Platform, which includes the OpenShell open-source secure runtime and the Sentry security framework. The significance is architectural: policy is set centrally in Control Tower but enforced at the point of execution, through the network and through an agent SDK, rather than relying on agents to police themselves.

That matters because the hardest governance problems are behavioral. Agents choose tools dynamically and pass work to other agents, which makes static permissions brittle. The infrastructure layer can also provide independent protection if an agent escapes its software boundary or its host is compromised.

The shadow fleet and the cost meter

The Control Tower expansion reaches beyond credentials. Shadow AI detection, built on discovery through more than 400 out-of-box connectors and network scanning, finds unapproved agents running on the network. Access mapping grounds authorization decisions in actual entitlements, and continuous observability tracks agent traces and behavior. According to G2, 38 percent of organizations with agents in production operate more than 100 of them, while only 27 percent maintain a central agent registry and just 14 percent have real-time observability into unusual behavior, costs, or risks.

Cost is becoming a governance problem in its own right. AI Cost Management, expected in the fourth quarter of 2026 as part of IT Asset Management, will bring frontier model, embedded SaaS AI, hyperscaler, and data center GPU costs into a single view and alert teams to potential budget overruns, with graduated responses escalating from alert to throttle to block. On pricing, ServiceNow described a hybrid model combining predictable seat-based licensing with "assists," its credit system for AI consumption.

Employees get to build, under guardrails

The same announcement introduced the third generation of ServiceNow's employee-centric AI platform. AI Agent Studio lets employees build AI agents through natural language, with governance controls embedded by default so IT keeps policy enforcement over employee-created agents. Now Intelligence delivers contextual, role-specific assistance across the Now Platform, including intelligent search, content generation, and task automation. The employee harness ships with the Yokohama platform release.

EmployeeWorks, the employee-facing AI front door that launched in February and now counts millions of users across its web app, Teams, and Slack in a hundred languages, gained persona-based productivity assistants that employees can tailor within administrator-defined guardrails, plus no-code productivity routines for repeatable background tasks. A Conversations API exposes the system to third-party apps and agents, including Claude and ChatGPT.

What to watch

AI Passport, shadow AI detection, and AI Cost Management are all expected in the fourth quarter of 2026. ServiceNow's own release cautions that delays, regulatory changes, and uncertain demand could alter what is delivered, so treat the timeline as a plan rather than a promise. For financial services, an Operational Resilience solution giving chief risk and security officers a real-time view of ungoverned assets, identities, and agents is available now.

The biggest open question is adoption beyond ServiceNow's ecosystem. In the near term, expect human approvers to remain the starting point, with autonomy granted gradually and an exit always available. The passport era has begun, but the borders are still being drawn.