How Do You Prove an AI Agent Is Really Who It Says It Is?

Nov

20

2026

Time to read

Read so far

Written by: 

Rishi Kaushal

Time to read

Written by: 

AI agent chip with a green checkmark on a digital circuit board

The next identity crisis may not be about proving that a person is human. It may be about proving that an AI agent is legitimate. AI agents are quickly evolving from assistants that generate answers into digital actors that can access sensitive data, call APIs, use enterprise applications, and take actions on behalf of people and organizations. As their autonomy grows, an uncomfortable identity question is moving to the center of enterprise security: How do you know an AI agent really is who it claims to be, and that it can be trusted to do what it's doing?

Key Takeaways:

  • Agent identity is essentially a trust problem. Organizations need to establish not only that an agent is legitimate, but what it is authorized to do and whether it can still be trusted in the moment.
  • The fraud landscape provides a warning about what happens when trust becomes too dependent on any one signal or any one point in time. Like fraud prevention and detection, agent identity needs the same philosophical shift towards continuous verifiable trust.
  • Identity is the control plane for continuous agent trust. The goal isn't simply to issue an agent a credential and declare it trusted, it is to create a continuous verifiable chain of trust. Because, as agents gain more autonomy, the consequences of misplaced trust grow with them and that is the identity challenge of the agentic era.

Agents Complicate Digital Identity

For decades, digital identity has largely centered on people. We establish who someone is, authenticate them when they return and determine what they are allowed to access and do.

Agents complicate that model. They are not people, but they can act with human-like agency. And increasingly, they operate with access to the same sensitive systems, data, and transactions.

When an agent requests access or attempts an action, “Who are you?” is only the beginning. Organizations also need to ask: Who created and owns this agent? What is it authorized to do? Is it running in the intended environment? Should this specific action be trusted now?

And major banks are learning the perils of personal agents like Meta’s Muse and OpenAI’s ChatGPT in real time. JP Morgan, Bank of America, and Wells Fargo have all detected unauthorized third party AI agents operating inside customer accounts and in some cases taking actions that appear independent of user instruction. While most US banks are moving in the direction of allowing read only personal agent access to customer accounts, the majority of Canada’s major banks are moving towards an outright block of AI agents from directly accessing customer accounts.

What Fraud Teaches Us About Agent Trust

The fraud landscape offers a useful warning for agent identity: something that looks legitimate is no longer sufficient proof that it should be trusted.

According to the 2027 Entrust Fraud Report, overall document fraud fell to 2.5% in 2026, yet the attacks that remain are becoming more sophisticated. Medium-sophistication document attacks rose from approximately 2% of attacks in 2025 to 39.4% in 2026, while the most sophisticated attacks increased 3.7 times year-over-year. At the same time, deepfake attacks increased 165% year-over-year, injection attack volume rose 224%, and reported social-engineering attempts increased 520%. Device-intelligence fraud detections also increased 42% year-over-year.

The lesson for agent identity is clear: do not trust a single signal. An agent may present a valid credential while operating outside its intended purpose. Its credential may be stolen, its runtime altered, its permissions abused, or its legitimate identity used to perform an illegitimate action.

Organizations need to move beyond “Is this credential authentic?” to “Should I trust this agent, in this environment, performing this action, right now?”

That decision should combine agent identity with contextual evidence, including the calling application, runtime environment, session, requested resource, current permissions, and the sensitivity of the action.

Another key takeaway from the Entrust Fraud Report is that attackers are increasingly targeting trust after it has already been established. In banking, account takeover now represents 52% of fraud, compared with 48% associated with new accounts. Payments and professional services show a similar pattern, with more fraud attributed to account takeover than new accounts. That's why organizations need to reinforce trust during events including authentication, re-verification, account recovery, and credential changes, rather than concentrating controls exclusively at onboarding.

Agent identity needs the same lifecycle approach. An agent that was trusted yesterday should not automatically be trusted forever. A high-risk transaction, privilege change, new tool, unusual behavior, or access to particularly sensitive information can become a new moment of truth at which trust needs to be reassessed.

Agent Identity Requires A Verifiable Chain of Trust

A trustworthy agent needs several layers of proof working together.

  1. Establish the agent’s identity

    Every agent should start with a unique machine identity backed by a cryptographic credential.

    That identity provides a foundation for distinguishing one agent from another and creating an accountable relationship between the agent and the organization, application, or individual responsible for it.

    But establishing identity is only the beginning.

  2. Prove possession

    Presenting a credential isn't enough if that credential can be copied, stolen, or replayed.

    Cryptographic proof can help establish that the entity requesting access actually controls the credential associated with the agent. That changes the question from simply, “Do you have an identity?" to, "Can you prove this identity is yours?"

  3. Verify agent provenance

    Organizations also need confidence in the provenance of the agent itself.

    That means connecting identity with signals about the agent's software, configuration, and operating environment. A trusted agent identity attached to an altered or compromised workload shouldn't automatically remain trusted.

    In other words, identity tells you which agent is calling. Provenance helps tell you whether it is still the agent you intended to trust.

  4. Separate agent identity from authority

    Knowing who an agent is doesn't mean it should be allowed to do everything.

    Agent identity must be connected to clearly defined authorization: which systems it can access, which information it can retrieve, and which actions it is permitted to perform.

    For consequential operations, organizations can also introduce stronger authorization requirements or human approval.

  5. Evaluate the action in context

    Even an authenticated, authorized agent may make a request that is inappropriate for the current session or risk level. The requested action must align with the agent’s identity, integrity, environment, and granted authority.

The resulting chain is simple:

Owner → Agent → Authority → Action

Each link should be verifiable, especially for high risk actions.

From Agent Authentication to Continuous Trust

The 2027 Entrust Fraud Report describes fraud as increasingly sophisticated, multichannel, and lifecycle-oriented. Its recommended defensive direction similarly moves toward layers of document verification, biometrics, device intelligence, re-verification, and behavioral monitoring across the identity lifecycle. Agent identity presents with a different technical problem, but the same continuous trust principle applies.

A continuous-trust model brings the necessary controls together:

  • Identity establishes which agent is acting.
  • Cryptography proves credential possession and protects agent identity integrity.
  • Provenance shows what software is running and where it came from.
  • Authorization defines what the agent may do.
  • Context determines whether the current action should be trusted.
  • Continuous verification determines whether trust should persist.
  • Auditability creates accountability for what happened.

The opportunity is to build those principles into agentic ecosystems now, before autonomous agents become deeply embedded across enterprise workflows.

Identity Is the Control Plane for Agentic Trust

The shift to agentic AI forces us to rethink what it means to prove identity.

For people, organizations have traditionally focused on establishing identity, authenticating that individual, and determining what they can access. Agents add another dimension: we need to establish not only who they are, but who stands behind them, what authority they have been given, whether they are operating as intended and whether their actions remain trustworthy.

The fraud landscape provides a warning about what happens when trust becomes too dependent on any one signal or at any one point in time. Agent identity needs the same philosophical shift.

The goal isn't simply to issue an agent a credential and declare it trusted. It is to create a verifiable chain of trust. Because, as agents gain more autonomy, the consequences of misplaced trust grow with them.

That is the identity challenge of the agentic era. And solving it may determine whether autonomous AI can become accountable AI.

Rishi Kaushal headshot
Rishi Kaushal
Chief Information Officer

Rishi Kaushal is Chief Information Officer at Entrust, where he leads the company’s global IT strategy, digital transformation initiatives, cloud modernization efforts, and enterprise technology operations. With more than 20 years of experience in IT leadership, Rishi has helped drive corporate growth through technology innovation, M&A integration, and business transformation. Prior to joining Entrust in 2019, he held leadership roles at HID Global, Toshiba America, and DAZ Systems (now part of Accenture).

View all of Rishi's Posts
Facebook