Cloud Computing and Data Centers

Who Owns the Cloud Resource Created by an AI Agent?

The article explains that AI agents may create cloud resources without a clear human owner, leaving organizations with ongoing costs and incomplete audit records. It proposes three practical controls: discovering machine identities, preventing deployments without a human owner, and setting environment lifetimes.

2026-10-08
4 min read
397 views
certi.news Editorial Team
Who Owns the Cloud Resource Created by an AI Agent?

When an AI agent creates a virtual network, databases, or a load balancer, the problem does not end when the task is complete. The resource may remain running, consuming budget and requiring maintenance, while its ownership tag points to a machine identity such as deploy-agent rather than to a person who can be contacted or held responsible for the resource.

The article presents this issue as a governance gap in cloud infrastructure: the execution identity that created the resource is not necessarily its actual owner. As the number of agents and the speed at which they create resources increase, traditional employee-based ownership mechanisms become insufficient.

Discovering Resources Owned by Agents

The first step is to inventory resources that have an owner tag, then compare the value of that tag with the organization's identity directory. In AWS and Azure, the tag can be matched against records of roles or native service principals. In Google Cloud, the article proposes verifying that the owner's value exists in the employee directory; if it does not, then it is not a human identity under the checking rule being used.

The appearance of a machine identity does not mean that the resource is illegitimate; the automation may have been designed to operate infrastructure on behalf of the organization. However, this check narrows the review scope to resources that actually need a human owner to be assigned.

What Changes in Practice?

Instead of automatically accepting the person or account that executed the deployment plan as the owner, the approval policy should verify the ownership tag itself. The proposed rule is that no deployment should pass if the taggable resource has no owner, if the owner's value is not in the form of a person's email address, or if it matches one of the known agent identities.

The article recommends keeping the list of agent identities in a JSON file that is passed to the policy engine. If the list is absent, the policy should fail closed rather than allow deployment without inspection. It also warns against using API keys with administrative privileges; instead, it proposes giving each agent a User-type key within a separate team and with limited permissions on the environment in which it operates, while avoiding shared tokens.

Setting Environment Lifetimes

Deployment policies alone are not enough to address old resources that agents have already created. The article therefore proposes using an environment expiration period, with a person creating the environment and setting its lifetime while the agent performs the deployment inside it. In this way, expiration warnings reach a human party capable of taking action instead of being sent to a machine identity that does not have a mailbox.

The article states that env zero warns the environment creator before automatically destroying it at two intervals: two days and two hours and thirty minutes. It also refers to a model at Axiom that automatically deletes an account if it is not claimed within 24 hours.

Why Does This News Matter?

The article links the growing scale of the problem to the increasing number of agents within organizations. According to Gravitee's 2026 survey, the average number of enterprise agents nearly doubled over four months, between December 2025 and April 2026, while more than one-third of organizations were operating more than 100 agents. Surveys conducted by Larridin also found an average of 47 agents in enterprise environments without an assigned owner.

Investigation becomes more difficult when agents use a shared service account; the log may prove that the resource was created, but it does not clarify which execution, request, or decision led to it. Therefore, assigning an explicit human owner addresses not only costs but also part of the accountability gap in audit records.

The open limitations are also clear: the success of these controls depends on the accuracy of identity directories, on agents not bypassing authorization policies, and on the existence of a consistent organizational definition of who actually owns the resource. The source itself is a guidance article sponsored by env zero, so its product-specific recommendations should be treated as proposed solutions rather than as the market's sole standard.

News source
The New Stack - Software Development
Open original source ↗
c
Author

certi.news Editorial Team

What you need to know

AI agents may create cloud resources without a clear human owner, leading to ongoing costs and incomplete audit logs. The article proposes discovering machine identities, preventing deployments without a human owner, and defining lifespans for cloud environments.

  • The execution identity that created the cloud resource is not necessarily its actual owner.
  • Ownership tagging should be checked against the organization's identity directory to identify resources associated with machine identities.
  • The policy should reject deployments if no human owner is present or if the owner value is a known agent identity.
  • Each agent should be given an identity and limited permissions, while administrative API keys and shared tokens should be avoided.
  • An environment expiration period helps send expiration warnings to a human before automatic destruction.
  • The effectiveness of these controls depends on the accuracy of identity directories, compliance with permission policies, and the organization's definition of ownership.

FAQ

Why are resources created by AI agents a problem?

Resources may remain running after a task ends, consuming budget and requiring maintenance without a clear human owner.

How can agent-owned resources be detected?

The owner tag is checked against the organization's identity directory; if it matches a machine identity or is not found in the employee directory, the resource is flagged for review.

What control is proposed before deploying a cloud resource?

The deployment should not proceed if the resource has no owner, if the owner value is not a person's email address, or if it matches a known agent identity.

What is the benefit of defining an environment lifespan?

Expiration warnings reach a person who can take action instead of being sent to a machine identity that has no mailbox.

In the same category

You may also like

View all news