OIDC-A and IETF AIMS — New Security Standards Replacing API Keys for AI Agents

The security practice of hardcoding API keys in code and manually renewing them upon expiration is no longer suitable for the era of AI agents. Because autonomous agents that make their own decisions need to browse the web and securely handle payments on our behalf, they require 'dynamic digital identities' rather than static passwords to verify themselves. Here is a clear breakdown of the core of the new agent-specific identity standards being rapidly proposed by global standards organizations.

SPIFFE and DID — How to Issue IDs for Agents

For agents to be trusted as independent members of the internet ecosystem rather than just simple automation tools, they must have verifiable identities that everyone can trust. The Agent Identity Management System (AIMS) standard, currently under discussion by the IETF, is a prime example. This standard treats agents as secure units of work running within a system and assigns them unique cryptographic addresses such as spiffe://회사명/agents/analyst. Just as we assign social security numbers to people, this creates a persistent digital identity for agents.

The W3C's agent identity registry standards complement this. Agents with long-term activity use web-based Decentralized Identifiers (DIDs), while ephemeral agents use one-time keys. This process leverages secure Ed25519 digital signatures and Verifiable Credentials (VC). As a result, receiving servers can transparently verify in real-time who the agent's true owner is and what permissions it has been delegated.

Thanks to these changes, developers no longer need to hardcode vulnerable text-based API keys into agent source code. Instead, we are building a cleaner and more robust internet ecosystem where agents present their own encrypted digital identities to communicate safely with other services.

OAuth AAP — Restricting an Agent's 'Scope of Action'

What happens if an authorized agent malfunctions, leading to unauthorized payments or the deletion of important data? To prevent such incidents and establish safe boundaries for agent actions, the IETF has proposed the Agent Authorization Profile (AAP) standard.

AAP functions by embedding five precise control mechanisms within the digital token used by the agent. This clearly defines the boundaries of what an agent can and cannot do.

With such sophisticated controls built directly into the token, developers no longer need to grant agents full access, but can restrict them securely to only what is necessary. This establishes the foundation for the principle of least privilege, which is essential in an era where agents think and act on their own.

OIDC-A — A Digital Baton Tracking 'Who Directed Which Agent'

When multiple agents collaborate, unexpected security holes can emerge. A task might be assumed to be directly requested by a user, when in fact authorization may have leaked during a chain of delegations between agents. To securely control such complex delegation, the OpenID Connect Agent profile (OIDC-A) has been introduced.

The core of OIDC-A is the agent_path information contained within the token. It is similar to a relay race where runners stamp the baton each time it is passed. For example, if a user wakes up a calendar management agent, which then calls a hotel booking agent, the token acting as the baton will record the identities of the user and each agent in sequence in real-time.

Thanks to this tracking method, the server receiving the final request can immediately verify that the command has passed through legitimate channels without being intercepted. This helps prevent agents from impersonating humans and creates a transparent audit trail to assign accountability if problems arise.

Microsoft Entra Agent ID — Identity Management in Real-World Services

While the standards discussed above are drafts actively being debated by global committees, some real-world technologies have already been applied to business. In May 2026, Microsoft officially launched Entra Agent ID, a security framework for managing and controlling AI agents. It defines agents not merely as software tools, but as independent 'non-human identities' that must be treated and managed on par with regular employees.

Entra Agent ID uses an intuitive three-tier structure to verify agent identity instead of complex manual permission settings.

Thanks to this, corporate security teams can now easily set agent-specific conditional access policies using management tools like Agent 365. They can transparently monitor which agents have entered the corporate network and what data they have read, and can immediately block suspicious activity. This is a prime example of how theoretical standard discussions are being implemented as robust security infrastructure in actual corporate cloud environments.

The Start of an API Key-Free Agent Ecosystem

Are you currently hardcoding API keys into the configuration files of your AI service under development? For future services utilizing agents, I recommend moving away from static keys that are prone to leaks and instead considering an encrypted identity system that verifies itself and manages connections securely.

The various international standards and Big Tech solutions are already pointing in one direction: a structure where agents act not just as automation tools, but as trusted partners collaborating safely with us in the internet world.