Published: 10 October 2026
The US National Institute of Standards and Technology is advancing work on identity and authorisation for software and artificial-intelligence agents, an area becoming urgent as autonomous systems gain access to code, business data and operational tools.
NIST’s National Cybersecurity Center of Excellence has summarised public feedback on its agent identity concept paper and selected enterprise software development as the first implementation use case. The planned demonstration will examine how agents can be identified, authenticated and authorised within the software-development lifecycle.
Why an agent needs an identity
Traditional access systems are built around people, service accounts and applications. AI agents complicate that model because they can plan work, call tools, create sub-agents and act with varying levels of human supervision.
If several agents share one user account or a broadly privileged service credential, an organisation may struggle to determine which system performed an action and why. That weakens accountability and makes incident investigation harder.
A distinct agent identity can support narrower permissions, clearer audit records and faster revocation. It also allows security teams to separate the authority of the human assigning a task from the authority granted to the agent executing it.
Context changes the security decision
Public comments collected by NIST indicate that agent governance may require richer evidence than conventional authentication. Access decisions may need to consider not only which agent is requesting permission, but also the task, tool, data source and operating context.
This is particularly important when an agent processes untrusted content. A malicious document, message or website can contain instructions intended to redirect the agent. Security policy may therefore need to tighten dynamically when prompt injection or other manipulation is suspected.
The problem also includes scale. Organisations may create large numbers of temporary agents for specialised tasks. Manually assigning and reviewing permissions for each one would be impractical, while permanent broad access would be dangerous.
Software development is the right first test
Development environments expose the challenge clearly. An agent may need to read a repository, run tests, open a change request or deploy code. Those actions carry very different levels of risk and should not rely on one blanket permission.
A standards-based demonstration can show how short-lived credentials, workload identity, human approval and detailed logging fit together. The goal should be useful autonomy with boundaries that are visible and enforceable.
NIST’s work signals a shift in the AI-security debate. The question is no longer only whether a model produces safe output. It is also whether an autonomous actor has the correct identity, the minimum necessary authority and a record that allows every consequential action to be traced.