For years, identity security has been based on a fairly straightforward premise: organisations need to know who, or what, is accessing their systems, what they are authorised to do, and whether that access remains appropriate. However, the rise of autonomous AI agents has changed the game, and tore-up the traditional model of role-based access controls and privileged access management.

AI agents are not simply another application or service account to add to the identity inventory. They can be granted access to enterprise systems, handle sensitive data, and increasingly make decisions or take actions on behalf of users. This combination of access and autonomy creates a new category of identity risk that organisations need to prepare for.

AI Agents Present a New Type of Identity

Traditional identity controls largely assume that access can be tied either to a person, whose behaviour is ultimately governed by human accountability, or to a machine identity carrying out a predetermined function.

AI agents sit somewhere between the two. An agent might be given a legitimate business objective – but determine for itself how to accomplish it with the tools and permissions available. Its behaviour may therefore be less predictable than that of a conventional service account.

This distinction matters. Authorising an AI agent to access an application is not necessarily the equivalent of giving a human employee the same permissions. The agent may be capable of completing tasks much faster, interacting with multiple systems in succession and taking actions at a scale that would be impossible for an individual user.

As a result, organisations need to look beyond whether an agent can access something and consider what it is capable of doing once access has been granted.

Identity Controls Need to Evolve

Many established identity platforms were designed before agentic AI became an enterprise consideration. Existing principles such as least privilege, strong authentication, and access review remain critical, but applying them effectively to autonomous identities requires additional controls.

Organisations need visibility of every AI agent operating across the environment, the systems and data each agent can access, the actions it is authorised to perform, and the identities or applications with which it can interact.

Identity security cannot end at the point of authentication. Organisations must be able to monitor what an AI agent actually does after it gains legitimate access, which makes behavioural monitoring increasingly important.

If an AI agent begins accessing unusual data, interacting with unexpected systems, or behaving differently from its established pattern, security teams need to identify and investigate that activity with speed. Detection and response therefore become closely intertwined with identity governance.

Governance is Just as Important

Perhaps the more difficult challenge is not technical but organisational. Human identities tend to exist within established accountability structures, where employees have managers, job descriptions and defined responsibilities. Their access can be reviewed against the role that they perform, while disciplinary and governance processes provide accountability when policies are not followed.

The lines, however, can become blurred when an AI agent takes action. Every agent should therefore have a clearly defined human or business owner who is accountable for its purpose, permissions and behaviour.

Organisations also need processes governing how agents are created, approved, monitored and modified. Questions such as who authorised an agent, why it requires particular access, and who is accountable if its behaviour creates a security incident, should be answered before deployment – not after something goes wrong.

This requires AI governance and identity governance to converge. AI agents cannot be treated purely as a technology issue if they are increasingly performing functions that influence business processes, sensitive information and security outcomes.

The Risk of Third-Party AI Agents

The issue becomes more complex when AI agents enter the organisation through third parties.

Businesses already depend on extensive ecosystems of suppliers, SaaS platforms, and technology partners. As those third parties embed agentic AI into their own services, organisations may increasingly find themselves exposed to AI identities that they did not create and do not directly control. This adds another dimension to traditional third-party cyber risk management.

Security teams need to understand not only which suppliers have access to organisational systems and data, but whether those suppliers are using AI agents, what those agents can access, and the controls – if any – that govern them.

Third-party access should require its own lifecycle. Each external account should have a named internal sponsor, an end date, and expiry by default unless the sponsor renews the access. This should be scoped to the systems that the work necessitates, and ensure that the supplier is contractually responsible for communicating when there are people changes within their business. Silence from the third-party should mean that access ends.

A Continuous Approach to Identity Security

As AI agents become embedded and used in everyday business processes, the organisations best positioned to adopt them safely will be those that treat identity as a continuous security challenge, rather than a one-time access decision.

AI agents may be a new category of identity, but the underlying objective remains familiar: ensure that every identity is known, appropriately governed, and continuously monitored throughout its lifecycle.  The difference in the age of agentic AI is that the user or system behind an access request may increasingly be capable of making decision of its own, and sometimes to the detriment of that organisation’s security.