Know Your Agent (KYA) is an emerging approach to establishing trust, accountability, and authorization for AI agents. While definitions vary across enterprise security and agentic commerce, the underlying principles are consistent: organizations need to establish which AI agent is acting, who is accountable for it, what authority it has been granted, and whether its actions remain within that authority.

Consider an AI-powered invoice processing agent that receives a supplier’s email requesting a change to the bank account associated with an upcoming payment. The agent’s credentials are valid, and the employee responsible for its deployment is legitimate. However, that employee only authorized the agent to match invoices against purchase orders, not to modify payment instructions.

If the agent interprets the supplier’s email as permission to change banking information, it could execute an unauthorized transaction despite being properly authenticated.

This scenario illustrates why identifying an AI agent is only the starting point. KYA aims to establish a verifiable relationship between an agent, its owner, and its delegated authority. When combined with continuous risk monitoring and policy enforcement, this context can help organizations detect unauthorized behavior and intervene before an agent’s actions create significant business or security exposure.

What is know your agent (KYA)?

Know Your Agent connects an AI agent’s identity to the human or organization responsible for its operation and the authority under which it acts. This may involve identifying the agent’s provider, recognizing a specific enterprise deployment, and documenting the permissions granted to that deployment.

Assigning a distinct identity to an AI agent makes its activity distinguishable from other software, agents, and human users. An ownership record establishes accountability, while a delegation record defines the tasks, resources, and actions the agent is authorized to perform.

Together, these elements provide the foundation for evaluating whether an agent is operating within its intended boundaries.

However, KYA should not be confused with a complete AI agent security solution. Establishing an agent’s identity and delegated authority provides essential context, but detecting misuse also requires visibility into its activities, risk analysis, and mechanisms for enforcing authorization policies.

Where does human identity fit?

Human identity plays a critical role in establishing and maintaining trust in AI agents. In many enterprise environments, agents operate on behalf of employees, administrators, or other authorized individuals. Consequently, the integrity of an agent’s delegated authority depends partly on the identity and permissions of the person granting it.

Identity-based attacks can undermine this relationship. An attacker who compromises an employee’s credentials may be able to register an unauthorized agent, modify an existing agent’s permissions, or approve actions that the legitimate account holder never intended.

Strong human identity verification helps establish that the individual granting authority is who they claim to be. However, verifying someone’s identity does not automatically establish that they approved a particular action.

For high-risk operations, organizations may need explicit, transaction-specific approval from an appropriately authorized individual. This distinction is particularly important when AI agents initiate financial transactions, modify access permissions, or interact with sensitive information.

KYA therefore depends on more than identifying the AI agent itself. It requires maintaining a trustworthy relationship between the AI agent, the human or organization accountable for it, and the specific authority granted.

What are the main KYA use cases?

KYA addresses different security and accountability challenges depending on whether an AI agent operates within an enterprise or interacts with external organizations.

Internal agents typically operate within an established identity and access management environment. External agents, such as supplier automation tools or AI shopping assistants, introduce additional challenges because organizations must evaluate identity and authorization evidence across trust boundaries.

Common KYA use cases include:

  • Employee task automation. An internal AI agent performs tasks using enterprise systems. KYA connects its activity to an approved deployment, the employee or team responsible for its operation, and the permissions granted to complete specific tasks.
  • Third-party business workflows. A supplier’s AI agent requests access to enterprise records. Recognizing the supplier establishes part of the trust relationship, but access must still be restricted to the resources and activities the enterprise has authorized.
  • Agentic commerce. An AI shopping agent makes purchases on behalf of a customer. Merchants must establish the agent’s origin, identify the customer it represents, and determine whether a transaction falls within the customer’s authorization.
  • Customer service and sensitive records. An AI agent retrieves account information or assists with a healthcare inquiry. The agent’s authority may vary depending on the customer, requested information, and specific task, even when the same agent serves multiple individuals.
  • Help desk and account recovery. An AI agent assists with credential resets or account recovery requests. Permission to collect information or initiate a workflow does not necessarily include authority to approve an account change or grant access.

Across these scenarios, the central challenge remains the same: establishing whether an identifiable AI agent is acting on behalf of an authorized party and within the permissions granted for a particular task.

Why can an authenticated AI agent still create risk?

Authentication and authorization serve different security purposes. Authentication establishes the identity associated with presented credentials, while authorization determines which resources and actions that identity is permitted to access.

A successful authentication event does not necessarily establish that the credentials are being used by their legitimate holder. More importantly, it does not guarantee that subsequent actions reflect the intentions of the person or organization responsible for the agent.

This distinction becomes particularly important when AI agents encounter untrusted information while performing legitimate tasks.

In a May 2025 demonstration, security researchers at Invariant Labs illustrated this vulnerability by embedding malicious instructions within content accessible through a GitHub integration. An AI agent encountered those instructions, retrieved information from a private repository, and disclosed it through a public pull request.

The demonstration highlighted how prompt injection can manipulate an AI agent into misusing legitimate access. The agent did not necessarily need to compromise credentials or bypass authentication controls. Instead, malicious instructions embedded in external content redirected its behavior.

The invoice processing example illustrates a similar problem. If an AI agent interprets a supplier’s email as authorization to modify banking information, untrusted content has influenced an action requiring separate approval.

These scenarios demonstrate why authentication, authorization, and behavioral risk detection must work together.

KYA can establish which AI agent is acting and what authority it has been granted. However, detecting attempts to exceed that authority requires additional capabilities, including activity monitoring, policy evaluation, and risk-based intervention.

For high-risk actions, organizations may also require explicit human approval that is tied to the specific transaction rather than relying solely on the permissions originally granted to the agent.

What happens when an AI agent’s owner or access changes?

An AI agent’s relationship with its owner can change after deployment. Employees may transfer to different roles, lose access to sensitive systems, or leave the organization entirely. These changes can affect whether an agent should retain its existing permissions.

For AI agents operating on behalf of employees, this creates a direct dependency between the human identity lifecycle and the agent’s delegated authority.

For example, an employee who transfers from finance to another department may no longer be authorized to access payment records. An AI agent previously deployed on that employee’s behalf should not automatically retain access to those records simply because its credentials remain valid.

Similarly, an employee’s departure raises important questions about whether their AI agents should continue operating, which permissions should be revoked, and who assumes responsibility for any remaining workflows.

Complicating matters further, AI agents frequently interact with multiple applications and services. Updating an ownership record in one system does not necessarily revoke credentials, authorization grants, or active sessions in another.

As a result, outdated permissions can persist even after the original business justification for granting them has disappeared.

KYA must therefore extend beyond initial agent registration and identification. Its effectiveness depends on maintaining accurate ownership, delegation, and authorization relationships throughout the agent’s lifecycle.

Organizations also need mechanisms to detect changes that invalidate existing trust assumptions and ensure that downstream systems enforce updated access decisions.

Which attributes can be used to identify an AI agent?

AI agent identity remains an evolving area of cybersecurity, particularly when it comes to determining which attributes can reliably distinguish individual agents and their actions across different systems.

The National Institute of Standards and Technology’s Draft Concept Paper on AI agent identity raises questions about which identity metadata is essential for tracing AI agent activity and whether those attributes should remain fixed or vary according to the task being performed.

An AI agent’s name or descriptive label may identify a general software capability without distinguishing the specific deployment or execution responsible for an action.

Several categories of attributes can help establish more reliable identification and accountability:

  • Deployment identifiers. Unique identifiers connect authenticated activity to a particular AI agent deployment. Information about the execution environment provides additional context about where and how the agent operates.
  • Ownership and accountability. Stable identifiers for responsible individuals or organizations establish accountability. Delegation records document the permissions granted, while approval records can connect sensitive actions to the individuals who authorized them.
  • Configuration and capabilities. AI agent versions, configuration details, connected tools, and available integrations help establish what an agent is capable of doing and whether those capabilities align with its intended purpose.
  • Observed activity. Execution identifiers, tool invocation records, resource access events, and destination information help correlate actions with specific agent runs. These signals provide evidence for determining whether observed behavior remains within authorized boundaries.

Emerging implementations are already distinguishing between an AI agent’s general identity and its individual deployments.

For example, Microsoft’s Agent Identity Blueprint Documentation describes distinct AI agent identities that can share configuration through a common template. This illustrates why identifying a family of agents is different from identifying the particular agent responsible for an action.

The next challenge is connecting these attributes into a verifiable and continuously maintained trust relationship.

Identifying an agent becomes significantly more valuable when its identity can be correlated with its current owner, delegated authority, configuration, and observed behavior.

This provides enterprise security teams with the context needed to evaluate whether an agent still represents the individual or organization that authorized its operation.

What’s next for KYA?

Know Your Agent is evolving from an approach focused primarily on identifying AI agents into a broader framework for establishing accountability, managing delegated authority, and addressing identity-related risks throughout an agent’s lifecycle.

Significant challenges remain. Organizations need reliable methods for identifying individual agents, verifying delegated authority across systems, and maintaining accurate authorization records as agents, tools, and human responsibilities change.

NIST’s AI Agent Standards Initiative is exploring identity infrastructure and interoperability approaches that could help address these challenges.

However, identifying an AI agent and validating its credentials are only the starting points. Even a properly authenticated AI agent can be manipulated into performing unauthorized actions, particularly when malicious instructions influence its behavior or previously granted permissions no longer reflect the intentions of its human owner.

Addressing these risks requires organizations to move beyond point-in-time authentication toward continuous identity threat detection and risk mitigation.

By connecting AI agent identities with verified human identities, delegated permissions, and real-time risk signals, enterprises can gain greater visibility into suspicious behavior and emerging identity threats. When integrated with appropriate policy enforcement mechanisms, these capabilities can support additional verification, transaction-specific human approval, or restrictions on actions that exceed established authorization boundaries.

Importantly, KYA provides the identity and accountability context needed to inform these decisions. Continuous monitoring, behavioral analysis, and authorization enforcement provide the complementary capabilities required to detect and respond to misuse.

Ultimately, securing AI agents requires more than knowing which agent is acting. It requires continuously establishing who is accountable for its actions, what authority it has been granted, and whether its behavior remains within those boundaries.