In October 2025 the OpenID Foundation published Identity Management for Agentic AI: The new frontier of authorization, authentication, and security for an AI agent world, a whitepaper from its Artificial Intelligence Identity Management Community Group. The Foundation’s announcement, dated 7 October 2025, describes the paper as addressing “how to securely authenticate and authorize these autonomous systems while maintaining proper governance and accountability.” The paper is also posted to arXiv as 2510.25819.
The paper has two halves. The first, “Immediate Solutions to Current Use Cases,” describes what existing standards already handle: OAuth 2.1, the Model Context Protocol, Client Initiated Backchannel Authentication (CIBA), enterprise SSO and SCIM provisioning, and externalized authorization. Its summary of that half is that today’s patterns “can work for synchronous agents using multiple tools across a single trust domain, but not asynchronous or multi-domain contexts.” The second half, “Future-looking Problems for Autonomous Agent Identity and Authorization,” sets out what remains open: recursive delegation, scope attenuation, revocation, registries, and human governance at scale.
This page is a mapping, not a review. It maps the paper’s best practices and named challenges onto this doctrine’s constructs, and records where the doctrine’s proposed conformance criteria sit alongside them. The paper does not cite this doctrine, and this mapping does not suggest otherwise. The paper is a year old at the time of writing; it is mapped here because it remains the most complete statement from an identity standards body of what agent identity requires, not because it is new.
What was read. The whitepaper PDF linked below was read in full as retrieved on 20 October 2026. It is dated October 2025, and its arXiv record lists a single version. The OpenID Foundation’s announcement page was read for the facts stated above. The paper’s references, including the protocol specifications it cites, were not read, and nothing on this page characterizes them beyond what the paper itself says. Where this page and the OpenID Foundation disagree about the Foundation’s work, the Foundation is right.
The paper’s best practices
The paper closes its first half with six “general-purpose best practices.” Left column is the paper’s wording, verbatim.
| Whitepaper best practice | Where it lands in this doctrine |
|---|---|
| Use standard protocols. “Implement open frameworks like OAuth 2.1 for authentication and SCIM for lifecycle management rather than custom authentication, authorization, and provisioning mechanisms.” | The doctrine does not prescribe protocols; its Planes say where a control lives, not which standard implements it. Read that way, these are candidate implementations for Plane 1, Agent Identity & Lifecycle. The SCIM line has its own mapping on this site, against the SCIM Agent Governance Extension. |
| Authenticate agent interactions. “Most agent-to-resource and agent-to-agent interactions should be authenticated. While some contexts don’t require authentication, high-security contexts should never allow anonymous access.” | Law 1, Agents Are Identities, Not Tools; for agent-to-agent interactions, Law 4, Trust Does Not Travel |
| Apply least privilege rigorously. “Agents should not be granted broad access to protected resources, and their permissions should be managed in a way that is consistent with the permissions of the users to whom they will output information.” | Least Agency, stated under Law 1: “Least Privilege limits access. Least Agency limits autonomy.” Enforced through mission-scoped access boundaries at Plane 1. |
| Automate agent lifecycle management. “Address complex events like agent provisioning and de-provisioning across multiple systems, and ownership transfer by profiling standards like SCIM, rather than tightly coupling agent management to the workflow of a single user.” | The agent registry at Plane 1: provisioning, rotation and revocation; across systems, Law 3, Governance Must Span Systems |
| Maintain clear audit trails for governance. “Log all authentication events, authorization decisions, and agent actions.” | The reconstructable record at Plane 4, Human Oversight, Audit & Traceability: for any governed action, who requested it, under what authority, which policy decided, what was executed, and what resulted |
| Design for interoperability. “Build adaptable identity systems that can evolve with emerging standards like IPSIE and future MCP specifications.” | Law 3, and the second Non-Negotiable: enforcement must be a fabric, not a fragment, spanning every system the agent touches |
Where the paper locates enforcement
The paper’s section on operationalizing agent identity recommends the externalization of authorization logic, separating the Policy Enforcement Point (PEP) from the Policy Decision Point (PDP), and states the reason directly: “While an agent itself is a decision point for its own operational logic, the infrastructure it interacts with requires a separate authorization decision point to govern its actions.”
That is the distinction Law 2, Enforce at Runtime rests on, between a limit the agent is asked to observe and a limit the governing system enforces. In this doctrine’s architecture it spans Plane 2, Execution & Tool Governance, which governs what the agent can invoke, and Plane 3, Policy & Compliance Engine, which governs whether that invocation is compliant given current context. The paper makes the same point about natural-language instructions: the user approves the intuitive instruction, “while the system enforces auditable, deterministic resource constraints, ensuring the agent remains bounded even if it misinterprets the user’s intent.”
The challenges it names
The paper’s executive summary lists the “critical future challenges” that its second half develops. The ones with a direct counterpart in this doctrine are below, quoted verbatim.
| Whitepaper challenge | Where it lands in this doctrine |
|---|---|
| “User impersonation by agents should be replaced by delegated authority. Currently, agents often act indistinguishably from users, creating accountability gaps and security risks. True delegation requires explicit ‘on-behalf-of’ flows where agents prove their delegated scope while remaining identifiable as distinct from the user they represent.” | Law 1: an agent is an operational identity, not a feature of its user. The delegated scope is Law 4 and Plane 5, Multi-Agent Trust & Delegation. |
| “Recursive delegation creates risks. Agents spawning sub-agents or communicating tasks to other agents create complex authorization chains without clear scope attenuation mechanisms.” | Law 4: every handoff is a trust boundary, and “the agent on the receiving end inherits the task, not the authority.” Delegation scope definition at Plane 5. |
| “Scalability problems exist in human oversight & user consent. Users will face thousands of authorization requests as agents proliferate, creating security risks from reflexive approval.” | Law 5, Humans Retain the Right to Intervene; escalation paths from automated enforcement to human decision-making at Plane 4. The action tiers proposed in RFC 001 address the same problem, below. |
| “Trustworthy autonomy lacks automated verification. Scaling beyond human-in-the-loop safety models requires new, programmatic methods to ensure an agent’s actions continuously align with its operational goals and constraints.” | Law 2 and Pillar 4, Runtime Enforcement; Behavioral Drift on the doctrine’s Threat Surface is what continuous alignment checks are for |
| “Multi-facet behavior of agents complicates identity. Technological advancements allow an agent to act on its own, requiring agents to have their own credentials, permissions, and audit trails.” | Law 1; Plane 1; Pillar 1, Agent Identity |
Revocation, de-provisioning and the blast radius
The paper calls revocation “a critical, and largely unsolved, problem” in delegated architectures: “If a user revokes the primary agent’s access, there is no clear, immediate mechanism to propagate that revocation down a chain of offline tokens that may have already been further delegated.” It distinguishes revocation from de-provisioning, “the permanent and complete removal of the agent’s identity and its associated entitlements,” and warns that a merely revoked identity “may retain its underlying registration and trust relationships, representing a dormant but persistent threat.”
In this doctrine, revocation is an intervention under Law 5, trust revocation propagation across interaction chains is a named capability of Plane 5, and provisioning, rotation and revocation sit in the agent registry at Plane 1. The paper’s reason for urgency is the doctrine’s Cascading Failure threat in its own words: “its ability to recursively delegate authority means a single compromised identity can trigger a cascading failure across an entire ecosystem of sub-agents.”
The paper also closes what it calls the “auditability gap”: “an API call made by an agent on a user’s behalf is often logged indistinguishably from an action taken directly by the user, creating a black hole for accountability and forensics.” Its remedy is a credential carrying distinct identifiers for the human principal and the agent actor, so that audit logs “record not only who authorized an action but also which specific agent instance performed it.” That is the first two questions of the record at Plane 4: who requested an action, and under what authority.
What this doctrine provides alongside it
The paper names its own open problem as the absence of automated verification, and asks standards bodies to “accelerate the development of protocols that formalize these new concepts.” Protocols answer how authority is carried. The doctrine’s contribution at an adjacent layer is a proposed set of criteria for what must be true of an action, however the authority is carried, for that action to count as governed, stated so that an implementation can be tested against them and can fail. RFC 001, The Conformance Layer proposes twenty-nine such criteria, mechanism-agnostic by its own rules. It is a proposal open for comment, not doctrine; the doctrine in force is v4.
Read against the paper’s concerns, the relevant RFC 001 criteria are:
| Whitepaper concern | RFC 001 criterion (proposed) |
|---|---|
| Agents identifiable as distinct from the user they represent; the auditability gap | C1.2: agent-originated actions are distinguishable from human-originated actions without inference. C1.1: every action an agent takes appears in the record, attributed to exactly one agent, in every system the agent touches. |
| Least privilege; de-provisioning and dormant identities | C1.3: an agent’s authority is granted per mission, expires with the mission, and does not persist across missions or into dormancy. C3.5: an action whose effect is not enumerated is denied, not permitted. |
| A separate authorization decision point governing the agent’s actions | C3.1: the rules governing an agent’s behavior are external to the agent, not resident in its prompt, its weights, or state it authors itself. C3.2: policy is evaluated during execution, at each governed action, not once at session start. C3.8: an action whose applicable criteria cannot be evaluated does not proceed. |
| Recursive delegation and scope attenuation | C5.1: trust is established at every handoff, whether delegation, orchestration, tool invocation or subagent spawn, and never carried across one. C5.2: authority is never inherited from an invoking agent; a subagent’s authority is granted independently, scoped to its own mission. |
| Revocation that propagates | C4.1: a revocation, override or halt by an external authority is honored before the agent’s next commit. C5.4: on detection of anomalous behavior, the governing system halts every downstream consumer of the affected agent’s output, not only the agent. |
| Consent fatigue; risk-based escalation to a human | C4.2: every governed action carries an assigned tier, determined before it is attempted. C4.3: a Tier 3 action cannot commit without a recorded human authorization event that is non-delegable to any agent. C4.4: what the human is shown is generated by the governing system from the action’s committed parameters, not by the agent. |
| Audit trails for authorization decisions and agent actions | C3.3: the basis of every policy decision is recorded at decision time, traceable to a specific rule. C4.7: a mission crossing multiple systems produces one reconstructable record, not fragments per system. |
On consent fatigue the two documents arrive at the same shape from different directions. The paper proposes risk-based dynamic authorization, in which “routine, low-risk actions are permitted automatically” and an anomalous request triggers out-of-band human approval. RFC 001 grades actions by reversibility: Tier 1 actions may execute autonomously, Tier 2 actions require corroboration by a second principal, and only Tier 3, the irreversible, requires a human. Both reserve human attention for the actions where it changes the outcome.
An identity architecture and a conformance layer answer different questions: the first says how an agent proves who it is and what it was delegated, the second says what must be true for a given action to count as governed. They are designed to be used together. An implementation built on the patterns the paper describes could be tested against RFC 001’s criteria, and RFC 001 is open for comment through 31 October 2026 for exactly that kind of test. The paper invites feedback through its community group; that invitation is the OpenID Foundation’s, and nothing on this page responds to it.
Primary sources
- OpenID Foundation, Identity Management for Agentic AI: The new frontier of authorization, authentication, and security for an AI agent world (whitepaper PDF, dated October 2025, read in full as retrieved 20 October 2026; the source for every whitepaper quotation on this page)
- OpenID Foundation, announcement of the whitepaper (7 October 2025)
- arXiv 2510.25819 (doi:10.48550/arXiv.2510.25819; single version listed as of 20 October 2026)
Mapped 20 October 2026 against the October 2025 whitepaper. If the AI Identity Management Community Group publishes a revision, this page will be revised to match it rather than defended.