How is artificial intelligence changing the rules of Identity and Access Management, and why must identity governance extend beyond people to include their AI agents?

Artificial Intelligence is rapidly evolving from a tool that supports employees into an active participant in business operations. AI agents can now read emails, process documents, interact with CRM systems, create requests, and perform actions across multiple enterprise platforms.

However, this new reality raises a critical question for organizations: Who controls AI agents’ access to corporate resources, and who is accountable for their actions?

This was the central theme of two presentations in the AI Track at Bulgarian Identity Conference 2026: “How NOT to Get Hacked by Your AI Agent“ by Martin Kuvandzhiev, CEO of encorp.ai, and “Governing AI Agents as Identities: From Visibility to Compliance“ by Dr. Sascha Kern of Nexis.

When an AI Agent Is Not Hacked, but Your Company Still Loses Data

Martin Kuvandzhiev presented a scenario illustrating one of the most significant security risks associated with AI agents.

Imagine an agent with access to your inbox, corporate Drive, and the ability to send emails. It receives a message containing a hidden instruction to forward the latest contracts to an external email address. The agent interprets the message as an instruction and executes the action.

No password has been stolen. No malware has been installed. No traditional security breach has occurred.

The agent simply did what its permissions allowed it to do.

This example highlights the real challenge: securing AI agents is not just about protecting language models. It is fundamentally about managing their identities and access rights.

Some of the key threats include indirect prompt injection attacks, compromised tools and MCP servers, the use of borrowed credentials, and interactions between agents where one agent may exploit the broader permissions of another.

One particularly dangerous combination involves three capabilities:

  • Access to sensitive corporate data.
  • Exposure to untrusted content.
  • The ability to send information outside the organization.

When these three capabilities come together, seemingly harmless text can become a mechanism for data exfiltration.

AI Agents Are the New Digital Employees – But Who Manages Them?

In his presentation, Dr. Sascha Kern explored another critical aspect of the challenge: AI agents are increasingly being created directly by business users, bypassing established access governance processes.

An employee can create an agent in Copilot Studio, connect it to a CRM system, a ticketing platform, or corporate document repositories, and share it with the entire team.

An entirely new digital actor has now entered the organization. It may:

  • Have access to sensitive corporate resources.
  • Be used by multiple employees.
  • Hold its own permissions or inherit access rights from other identities.
  • Continue operating even after the person who created it has left the company.

The problem is that traditional Identity Governance and Administration (IGA) processes were designed around human identities. They assume that every user has an accountable owner, formally requested and approved access, periodic access reviews, and a clearly defined lifecycle.

With AI agents, these assumptions often no longer hold true.

Who Actually Determines an AI Agent’s Access Rights?

One of the key topics in Dr. Kern’s presentation was the distinction between the different ways AI agents obtain access to corporate systems.

  1. User Credentials (Delegated Access)

The agent acts on behalf of a specific user, using their permissions, limited by the scopes granted to the agent. This creates a dependency on the employee’s existing access rights and can expose information that the user technically has permission to access.

  1. Maker Credentials (Fixed Connections)

The agent uses a preconfigured connection based on its creator’s account or a service account. As a result, everyone using the agent may potentially gain access to resources available to that account. This creates a risk of privilege escalation without formal approval.

  1. Service Principal (Application Permissions)

The agent operates through its own application registration and permissions. While this allows independent access, it can also introduce excessive privileges, particularly when permissions are granted across the entire organization.

The situation becomes even more complex when different permission inheritance mechanisms are involved, including those managed through agent blueprints, roles, and policies in Microsoft Entra.

Therefore, knowing what permissions an agent has is not enough. Organizations must also understand where those permissions originate, who can use the agent, and what access each user effectively gains through it.

When a Person and Their AI Agent Bypass Security Controls Together

Another particularly interesting topic was Cross-Identity Segregation of Duties (SoD).

Consider an employee in the finance department who is authorized to release payments. Their AI agent, meanwhile, has permission to create and modify supplier records, including bank account details.

Individually, both identities may successfully pass standard segregation-of-duties checks. However, when the employee can invoke the agent, the combination of their permissions allows them to change a supplier’s bank details and subsequently release the payment.

This represents a significant shift in how organizations need to assess risk.

It is no longer sufficient to analyze employee and AI agent permissions separately. Organizations must understand what they can do together.

From Visibility to Governance and Compliance

According to Dr. Kern, effective AI agent governance should follow three interconnected stages:

Visibility

Organizations need a complete and up-to-date inventory of their AI agents, including:

  • Who created each agent and who is accountable for it.
  • Which systems and data it can access.
  • What permissions it uses and where they originate.
  • Which employees and groups can invoke it.
  • What risks emerge when its permissions are combined with those of its users.

Governance

The next step is to extend established identity governance processes to AI agents. This includes assigning accountable owners, conducting periodic access reviews and recertification, tracking changes, and automating lifecycle management.

Compliance & Security

Once visibility and governance are established, organizations can generate the necessary documentation and audit evidence, demonstrating control over access in the context of regulatory frameworks such as DORA, NIS2, and the EU AI Act.

Five Practical Steps Towards More Secure AI Agents

Both presentations highlighted practical measures organizations can implement before deploying their next AI agent:

  1. Create an inventory of all AI agents – including those developed by individual employees and integrated through external tools.
  2. Assign an owner to every agent – a named individual accountable for its access and use.
  3. Apply the principle of least privilege – start with read-only access and introduce write permissions only when necessary.
  4. Introduce human approval for critical actions – sending information, making payments, and deleting data should remain subject to appropriate human oversight.
  5. Ensure auditing, periodic reviews, and an immediate shutdown mechanism – every agent should have traceable actions and a mechanism to revoke its access.

The Future of Identity Management Includes Artificial Intelligence

The growing adoption of AI agents is changing not only how we work but also the very concept of digital identity.

AI agents are no longer just software tools. They are active participants with access to enterprise systems, permissions to execute actions, and the ability to influence business processes.

The main challenge for organizations is therefore not simply to adopt AI, but to do so while ensuring security, accountability, and control.

As the two presentations at Bulgarian Identity Conference 2026 clearly demonstrated, this does not require reinventing identity governance. It requires extending existing IAM and IGA principles to encompass this rapidly growing digital workforce.

Every AI agent needs its own identity, an accountable owner, clearly defined permissions, and an immediate shutdown mechanism.

Because in the age of AI, security starts with identity.


0 Comments

Leave a Reply

Avatar placeholder

Your email address will not be published. Required fields are marked *