Computer-Use Model

Expanded Definition

A computer-use model is an AI system that perceives a user interface, interprets screen state, and carries out actions through mouse, keyboard, or other interface controls. In NHI operations, it is not just a conversational assistant because it can enter the control path and affect production systems, credentials, or approvals.

Definitions vary across vendors on where a computer-use model ends and a broader AI agent begins, but the governance issue is consistent: once the model can click, type, submit, or approve, it has execution authority that must be scoped like any other privileged actor. NHI Management Group treats this as a control-plane concern, not only an AI safety concern, because the model may handle secrets, create change events, or trigger workflows that impact identity posture. For a broader NHI governance lens, see the Ultimate Guide to NHIs and the NIST Cybersecurity Framework 2.0 for control-oriented risk framing.

The most common misapplication is treating the model as read-only automation, which occurs when teams allow interface access without defining privilege boundaries, approval steps, or rollback conditions.

Examples and Use Cases

Implementing computer-use models rigorously often introduces latency and operational friction, requiring organisations to weigh faster task completion against tighter supervision, deterministic outputs, and stronger auditability.

• Resetting a locked service account by navigating an internal admin portal, where the model can fill forms but must not self-authorise credential changes.
• Updating metadata for API keys in a secrets workflow, with every action recorded because the model is touching identity records rather than drafting text.
• Processing helpdesk remediation steps after a secrets leak, guided by the patterns described in the Ultimate Guide to NHIs, while a human retains final approval for destructive actions.
• Operating a browser-based cloud console to rotate access tokens, using the same access governance discipline reflected in NIST Cybersecurity Framework 2.0.
• Triaging an identity alert by opening dashboards, copying evidence, and preparing a case for analysts, but not changing policy settings automatically.

In practice, these workflows are safest when the model has only the narrow interface scope needed for one task, with monitored sessions and version-pinned behavior. That is especially important when the model is chained to NHI workflows, because a small UI mistake can become a broad access event.

Why It Matters in NHI Security

Computer-use models matter because they can blur the line between advice and action. Once a model can operate internal tools, it may interact with secrets managers, admin consoles, ticketing systems, and approval flows. If its permissions are broader than intended, a prompt injection, UI change, or model drift can turn ordinary assistance into an identity incident. The risk is amplified when teams fail to log interface actions with the same rigor used for service accounts and privileged automation.

NHI Management Group research shows that 97% of NHIs carry excessive privileges, increasing unauthorised access and broadening the attack surface, which is a strong warning for any AI system given direct operational reach. The same body of research also shows that only 5.7% of organisations have full visibility into their service accounts, a gap that becomes more dangerous when an AI system is acting through those accounts or adjacent consoles. See the Ultimate Guide to NHIs for the underlying NHI governance context.

Organisations typically encounter the consequences only after a mistaken click, an unauthorized submission, or a compromised workflow, at which point computer-use model controls become operationally unavoidable to address.

Related resources from NHI Mgmt Group

• How should security teams govern AI agents that use Model Context Protocol?
• Should organisations use just-in-time access for AI model operations?
• How should security teams govern AI use when the same model creates different risk in different contexts?
• Should healthcare teams use the same zero trust model for AI agents and service accounts?