AI-first security will fail without identity control

AI adoption is moving faster than most identity programmes can safely support. As organisations move beyond pilots and begin deploying agents, copilots and automated workflows, they are exposing a control problem that was already there: too many identities, too much standing access and too little evidence of who or what can reach critical systems and data.

 

AI is already being connected to everything from collaboration platforms and customer data to operational systems and security tools. If those connections inherit excessive permissions, outdated entitlements or poorly governed shared spaces, AI can turn weak identity governance into an enterprise-wide access risk.

 

A well-governed AI service can only act within the limits that are set for it. An agentic system is different because it can reason across tools, access data sources, trigger workflows and operate through delegated permissions. If those permissions are poorly understood, the agent does not need to break the rules to create risk. It only needs to follow the access path the organisation has already left open.

 

Enterprise environments collect these risks over time. How often, for example, does a supplier get added to a shared workspace for a project, only for the work to finish while the folder and access remain in place? Connect an AI agent or copilot-style tool into that environment, and the issue is no longer just an old permission. It becomes part of the data and workflow context that AI can search or act on.

 

The difference now is that AI applications, autonomous agents and other non-human identities can inherit the same excessive access. That changes the problem from user administration to enterprise control: what exists, who owns it, what it can access and how quickly risk can be reduced when something changes.

 

AI is turning identity sprawl into a board-level risk

Identity now extends far beyond employees. It includes partners, contractors and guests, alongside service accounts, applications, devices, workloads and AI agents. That expansion is outpacing many governance models. IDC predicts AI agent use will surge tenfold by 2027. Palo Alto Networks, meanwhile, estimates there are 109 machine identities for every person in the average enterprise.

 

This rise of AI and non-human identities requires the same scrutiny, ownership and policy control as human users. Every identity that can access data, trigger a workflow or connect to another system needs a lifecycle, an owner and a clear access boundary.

 

A common weakness is that many non-human identities are created for a specific task, integration or automation, then forgotten when the owner changes role or leaves the business. Over time, this creates a complex web of accounts and permissions, along with secrets and dependencies that few teams can fully explain.

 

AI is making identity abuse easier to scale

Criminals target identities because a valid identity can look like legitimate activity. AI gives them new ways to obtain, abuse and disguise that access, from more convincing phishing and deepfake social engineering to automated reconnaissance and faster lateral movement once credentials or tokens are compromised.

 

Microsoft’s Secure Access in the Age of AI report found that organisations are still dealing with fragmented identity and network access controls as GenAI and AI agents expand the number of identities, access points and potential vulnerabilities across enterprise environments.

 

The governance implications extend beyond the security team. Customers, insurers, regulators and auditors increasingly want evidence that access to sensitive systems and data is understood, controlled and reviewed. With the EU’s AI Act increasing scrutiny on AI governance, organisations need to show how AI systems and machine identities are governed, what data they can reach and how access decisions are recorded.

 

Governance for AI-era identity

AI-era identity governance must work in live environments, not just in policy documents. The most valuable systems, data stores, collaboration spaces and automated workflows need named owners, clear access policies and review cycles that security teams can evidence.

 

Connected devices, software bots and automated processes need the same discipline as user accounts. Each should have an owner and an approved purpose, as well as documented access, secure authentication and a clearly-defined process for managing change or removal.

 

Privilege now extends beyond human users

Zero Trust remains essential, but it has to be applied to identities and workloads, as well as the applications and agents operating across the environment. Access should be conditional, risk-based and continuously reviewed, with privileged roles governed through just-in-time elevation, strong authentication and SOC visibility.

 

Privileged access should be time-bound, approved and reviewed. IoT devices, service accounts, application permissions and AI agents also need active lifecycle management, including joiner, mover and leaver processes for the systems, owners and automations they depend on.

 

Data protection has to be tied directly to identity control. Security teams need to know where sensitive data sits, who and what can access it, and whether that access is still justified. They also need controls that reduce the risk of sensitive information being exposed through generative AI services, third-party connectors or poorly governed prompts and outputs.

 

Bring identity control into the AI operating model

Identity governance, access management, data protection and security operations now need to operate as one control model. Security teams need visibility across people, apps, devices, workloads and agents, with enough context to see where access is excessive, where ownership is unclear and where risk is increasing.

 

AI adoption will continue to accelerate, but security teams cannot govern what they cannot see, explain or control. Organisations that treat identity as part of the AI operating model will have a stronger basis for secure adoption, regulatory evidence and blast-radius reduction when something goes wrong. Delayed identity governance leaves AI systems to inherit, accelerate and expose access decisions that were never designed for autonomous use.

 

---

 

Steven Jones is Chief Technology Officer at Kocho

 

Main image courtesy of iStockPhoto.com and WANAN YOSSINGKUM