When a person asks an AI assistant to organise a file, it may need permission to open a folder. When that same assistant can send messages, change records or make purchases, the question becomes much more serious: exactly whose authority is it using, and where does that authority end? The problem of AI agent identity has prompted researchers and standards bodies to examine how to identify agents as distinct actors rather than letting them behave as unrestricted extensions of a user’s account.
The US National Institute of Standards and Technology has published a concept paper on software and AI agent identity and authorisation. It explores how existing security approaches might apply when software agents act on behalf of people and organisations. The paper is a proposal for work and discussion, not a universal rule already enforced across AI products.
Why AI agent identity matters more than a clever prompt
A human employee often has access to several services because they carry responsibility for multiple aspects of a job. If an AI assistant receives all those privileges to complete one narrow task, a mistake can have much wider consequences. An agent asked to find a delivery update should not automatically be able to cancel orders, change payment details or browse unrelated personnel files.
The safest pattern is to create a separate identity for the agent and give it only the permissions required for a defined purpose. Security professionals call this least privilege. It is valuable even when a system is trusted, because reliable software can still encounter unexpected or malicious inputs. Restricting authority limits how far a problem can spread.
In a workplace, the agent’s identity should make it possible to tell its actions apart from those of the person who requested the work. Otherwise, an audit may show that a manager’s account changed a record without revealing whether the manager clicked the button or an automated system did so on their behalf. That distinction matters for correcting errors and assigning responsibility.
Instructions are not the same as authorisation
A model may read words on a web page, inside an email or in a downloaded document. Some of those words may be instructions intended for a human reader. Others may be planted by an attacker to influence the model. A secure system should not let outside text create new permissions simply because the model encounters it during an authorised task.
Consider a supplier document that claims a payment should be sent to a different account. Even if the request looks plausible, the assistant should not treat the statement as authority to change banking records. The real organisation needs a separate approval process. AI agents can make this risk more subtle because they may interpret a malicious passage as relevant to the user’s original request.
NIST’s AI Agent Standards Initiative includes work on agent security, interoperability and identity. It reflects recognition that authentication and permission controls designed for traditional software need careful application as AI becomes more autonomous. Existing cybersecurity principles remain important; the new challenge is connecting them to changing, model-directed actions.
A permission should have a purpose and an expiry
An agent that needs access for one task should not necessarily keep that access indefinitely. Short-lived credentials can reduce the impact if an account or tool is compromised. A clear purpose also helps administrators judge whether a proposed action belongs to the original request.
For example, an assistant preparing a travel itinerary may be allowed to read calendar availability and compare transport options. Booking a ticket introduces payment, cancellation terms and potentially sensitive identity information. A sensible workflow separates research from the final purchase and requires explicit approval for the consequential step.
The same principle applies to a small business using AI for bookkeeping. Matching transactions and drafting reports can be useful, but changing supplier bank details or making payments should require a stronger process. A system that works faster than a person should not be allowed to bypass established financial controls in the name of convenience.
How to reconstruct what happened
Identity is only part of the solution. Organisations need logs showing what information an agent accessed, what tool it used, whether a person approved an action and how the result was returned. Without a dependable record, a business may struggle to explain a mistake to a customer or regulator.
An AI agent’s own explanation is not sufficient evidence of every step it took. Models can summarise incompletely and may not know why another connected service behaved in a particular way. Surrounding systems should therefore record observable actions independently wherever practical. Those records need protection too, since they may contain confidential information about customers and staff.
LiveAIWire’s coverage of AI agents and corporate insider risk discusses the need for monitoring when software can reach sensitive systems. Identity and authorisation make that monitoring more meaningful by connecting an action to a specific agent and the authority it was given.
Why agent autonomy is a matter of degree
Not all agents require the same controls. A system that summarises public documents has different risks from one that edits payroll records. It makes sense to match the degree of autonomy to the task. Read-only access may be enough for research, while reversible changes may be appropriate for low-stakes administration. Irreversible or high-value actions deserve the strongest approval requirements.
The system should also know when to stop. If a tool fails or produces conflicting information, repeated attempts may make matters worse. The ability to report uncertainty and ask a person for a decision is a useful feature. More autonomy is not automatically better if it removes opportunities to prevent an expensive mistake.
LiveAIWire has examined agents working with email and calendars, which illustrates both the convenience and the sensitivity of ordinary integrations. Sending one message to the wrong person can disclose information or create a misunderstanding even if no money changes hands.
What a sensible business policy looks like
A company adopting an AI agent should begin with a specific job description. State what the agent may read, what it may write, what it must never do and who accepts responsibility for the outcome. Give it a separate identity if the technology permits, and avoid reusing broad human credentials. Test the workflow with deliberately awkward examples before allowing it near live records.
Review logs and permissions when the task changes. An agent originally built to answer customer questions may gradually acquire access to invoices, refunds and account management. That growth should trigger a fresh risk assessment rather than happen invisibly through convenient new integrations.
Clear communication matters as well. Staff should know when work is being performed by software and how to challenge a result. Customers should not be led to believe that a human personally reviewed every action when the process was automated. Transparency allows the organisation to explain decisions and fix problems without relying on vague assurances about AI quality.
The standards question is still open
NIST’s work is important because companies may eventually need agents built by different vendors to operate under comparable rules. A shared language for identity, delegation and evidence could make that easier. But publication of a concept paper is an early step. Interoperable standards, implementation guidance and real-world tests will determine whether the ideas become practical safeguards.
The immediate principle is already clear: a request to do a task is not permission to do anything a connected account can do. Treating an AI agent as a separately accountable actor, with narrowly defined authority and visible actions, offers a more trustworthy foundation than giving clever software the keys to everything.
Permission can be withdrawn as well as granted
Delegation should have an end point. If an employee leaves a company, a customer revokes consent or a project concludes, an AI agent should not continue acting under credentials that were previously available. Businesses need a process for finding every connected agent, reviewing its scope and disabling access without disrupting unrelated systems. Short-lived authorisation is helpful, but it is not a substitute for an up-to-date inventory of agents and their permitted functions.
The same principle affects consumer services. An assistant may be allowed to investigate a flight, but a buyer could change their mind before making a booking. A system needs to distinguish a past request to research from current permission to spend. LiveAIWire has examined the rules that still apply when AI shopping agents act for customers. Those obligations do not disappear simply because software rather than a person fills in the form.
Good security design makes the authorised scope visible to the user and understandable to an auditor. People should be able to ask what the agent can do, when its authority expires and how to reverse a mistake. If the answer is unclear, the safest decision is to restrict access until the purpose has been defined properly.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.
