Internet Engineering Task Force K. Pandey Internet-Draft TEKsystems, AI & ML Practice Intended status: Informational 17 November 2026 Expires: 21 May 2027 Session-Scoped Authorization for AI Agents: A Per-Tool-Call Authorization Profile Building on WIMSE and OAuth draft-pandey-agent-identity-00 Abstract Autonomous AI agents act with privileges: they call tools, read and write data, and operate services. In most current deployments they do so under shared service accounts, with permissions broader than any single task requires and without an audit trail that binds an action to the agent that took it and the human on whose behalf it acted. This document specifies a session-scoped authorization profile for AI agents, building on the WIMSE architecture and the OAuth 2.0 protocol family. Each agent is a first-class IAM principal with its own credentials. Each session is bound to an authenticated human sponsor through an authenticated delegation chain. Authorization is evaluated at every tool call, with effective permissions computed as the intersection -- never the union -- of the agent principal's permissions, the sponsor's permissions, and the session scope. Actions that cannot be evaluated are denied (fail closed), defined classes of irreversible actions require sponsor approval, and every evaluated tool call produces a tamper-evident audit receipt. The document defines the model, states requirements for implementations, maps the model onto existing OAuth and WIMSE mechanisms, and gives full security considerations. It defines no new protocol and requests no IANA actions. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 21 May 2027. Pandey Expires 21 May 2027 [Page 1] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License. Table of Contents 1. Introduction .................................................. 2 1.1. Problem Statement ........................................ 3 1.2. Scope and Relationship to Existing Work .................. 4 1.3. Requirements Language .................................... 5 2. Terminology ................................................... 5 3. Architecture and Roles ........................................ 7 4. Session-Scoped Authorization Model ............................ 8 4.1. Agents as First-Class Principals ......................... 8 4.2. Sponsorship and Delegation Chains ........................ 8 4.3. Session-Scoped Credentials ............................... 9 4.4. Per-Tool-Call Authorization and Permission Intersection .. 9 4.5. Dispositions and Fail-Closed Behavior ................... 10 4.6. Human Approval for Irreversible Actions ................. 11 4.7. Revocation .............................................. 11 4.8. Audit Receipts .......................................... 12 5. Protocol and Deployment Considerations ....................... 12 5.1. Token Contents and Identifiers .......................... 12 5.2. Mapping to OAuth and WIMSE Mechanisms ................... 13 5.3. Cost Attribution ........................................ 13 5.4. Deployment Experience and Open Issues ................... 14 6. Security Considerations ...................................... 14 6.1. Prompt Injection as Privilege Escalation ................ 15 6.2. Tool Squatting and Tool Substitution .................... 15 6.3. Confused Deputy ......................................... 15 6.4. Token Theft, Replay, and Long-Lived Credentials ......... 15 6.5. Delegation Chain Amplification and Splicing ............. 16 6.6. Audit Record Integrity and Privacy ...................... 16 6.7. Fail-Closed Denial of Service ........................... 16 6.8. Sponsor Compromise and Sponsor Authentication ........... 17 7. IANA Considerations .......................................... 17 8. References ................................................... 17 8.1. Normative References .................................... 17 8.2. Informative References .................................. 18 Author's Address ................................................. 20 1. Introduction AI agents -- software systems that use a language model to select and invoke tools across multiple steps toward a goal -- now perform Pandey Expires 21 May 2027 [Page 2] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 privileged operations in production environments: querying and modifying databases, invoking cloud APIs, moving files, sending messages, and operating infrastructure. The identity and authorization practices applied to these systems have not kept pace with that fact. The prevailing deployment pattern runs agents under shared service accounts: one credential, held by every instance of an agent (and often by several different agents), carrying the union of every permission any of them might need. That pattern fails in three specific ways. First, least privilege becomes unenforceable: any action permitted to the shared account is permitted to every agent running under it, so the effective authority of the system is the union of all tasks' needs rather than the need of the task at hand. Second, the blast radius of a compromised or misdirected agent is unbounded by task: an agent induced to misuse its access -- the realistic failure mode for language-driven systems is manipulation of the agent's inputs, not compromise of its code -- inherits the whole account. Third, actions become unattributable: when ten agent instances write through one credential, the audit log records what happened but not which agent acted, on whose behalf, or in pursuit of which task. Attribution is not a compliance nicety; it is the precondition for debugging, incident response, and access review. This document takes the position that the identity plane already has the right abstraction and that agents should use it: the agent is an IAM principal, with its own authentication, authorization boundary, and audit trail, subject to the same least-privilege discipline as any human or workload identity, plus controls for the failure modes specific to autonomous, language-driven actors. It specifies an authorization profile -- a set of requirements on how identities, sessions, credentials, and enforcement points behave -- rather than a new protocol. The profile is developed from the identity architecture described in [PANDEY-IAM] and is stated in terms of the WIMSE architecture [I-D.ietf-wimse-arch] and the OAuth 2.0 family [RFC6749]. 1.1. Problem Statement Existing mechanisms each cover part of the problem: o Workload identity systems, including the WIMSE architecture [I-D.ietf-wimse-arch], establish which workload holds which credential. They do not, by design, express what authority the workload holds, how that authority derives from a human originator, or whether onward delegation is permitted. o The AI Identity Management System document [I-D.ietf-wimse-aims] composes WIMSE workload identity with the OAuth 2.0 family to identify AI agents, including representing the acting agent and the party on whose behalf it acts. It addresses identification and authentication of agents; it does not specify a session scope model, per-action authorization semantics, revocation behavior for agent sessions, or the contents of audit records. Pandey Expires 21 May 2027 [Page 3] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 o The WIMSE applicability analysis for AI agents [I-D.ni-wimse-ai-agent-identity] maps agent scenarios onto WIMSE mechanisms, including owner-bound agent credentials. Session- scoped authorization, delegation-chain semantics, revocation, and audit records are outside its scope. o OAuth 2.0 [RFC6749] and its extensions provide delegation [RFC8693], sender-constrained tokens [RFC8705] [RFC9449], revocation [RFC7009], and introspection [RFC7662]. None of them defines how these mechanisms compose for an actor whose individual actions are chosen autonomously at runtime: in OAuth, the set of actions a client may take is fixed by the scopes granted at issuance time, whereas an agent selects each tool call -- its target and its arguments -- during the session, in response to content that may be adversarial. The gap this document addresses is therefore the authorization-time layer: what authority an identified agent holds during a session, how that authority is bounded by the human who originated the session, how each individual tool call is evaluated against that bound, what happens when a call cannot be evaluated, how authority is revoked, and what record each decision leaves behind. The NIST National Cybersecurity Center of Excellence has identified the same set of concerns -- identification, authorization, access delegation, and logging -- as the focus areas for software and AI agent identity [NIST-NCCOE-AI-ID]. Commercial identity platforms are converging on per-agent identities with accountable human sponsors: Amazon Bedrock AgentCore Identity issues workload identities per agent and supports on-behalf-of token flows [AGENTCORE], and Microsoft Entra Agent ID creates agent identities from blueprints, each with a designated sponsor accountable for it [ENTRA-AGENT-ID]. What those identities do at authorization time is the subject of this profile. 1.2. Scope and Relationship to Existing Work This document is an Informational profile. It: o defines a session-scoped authorization model for AI agents (Section 4); o states requirements, using the requirement keywords of BCP 14, for agent principals, sponsors, authorization servers, and policy enforcement points that implement the model; o maps the model onto existing OAuth 2.0 and WIMSE mechanisms (Section 5.2), in the composition style of [I-D.ietf-wimse-aims]: no new protocol elements are introduced; o defines the required contents of the per-tool-call audit receipt (Section 4.8). This document does not define a wire protocol, a token format, or a policy language. It does not address how an agent's identity is bootstrapped in adversarial environments, model-level safety, or Pandey Expires 21 May 2027 [Page 4] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 data-plane threats such as a legitimate tool returning poisoned data that an agent then acts on within its authority (see Section 6). Related mechanism work includes Transaction Tokens [I-D.ietf-oauth-transaction-tokens], which propagate immutable transaction context within a trust domain and can carry the session context this profile relies on, and [I-D.asor-wimse-agent-delegation-chain], which defines a token profile for verifiable, monotonically attenuated delegation across agent chains; the requirements in Section 4.2 are stated so that mechanisms of that kind can satisfy them. The SAGA architecture [SAGA] proposes a provider registry and cryptographic access-control tokens governing inter-agent communication; it addresses a complementary problem (which agents may communicate) to this document's (what an identified agent may do, per action, in a session). Risk and threat context for generative AI systems generally is given in [NIST-AI-600-1] and [OWASP-LLM-TOP10]. This is an individual submission. It has not been adopted by any working group, and nothing in it should be read as implying IETF consensus. Given the subject matter, comments may also be of interest to the WIMSE working group. 1.3. Requirements Language The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. 2. Terminology This document uses the terms "authorization server", "client", "resource server", and "access token" as defined by [RFC6749], the delegation and actor terminology of [RFC8693], and the workload terminology of [I-D.ietf-wimse-arch]. In addition: AI Agent: A software system that uses a language model or similar planner to select and invoke tools autonomously, over multiple steps, toward a goal, without per-step human initiation of each action. Tool: An invocable capability exposed to an agent -- an API operation, function, or service action -- whose execution can read or modify state outside the agent itself. Tool Call: A single attempted invocation of a tool by an agent, comprising the tool target and the arguments presented for that invocation. Agent Principal: The IAM identity assigned to an agent under which its actions are Pandey Expires 21 May 2027 [Page 5] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 authenticated and authorized. An agent principal is a first-class principal: it is provisioned, scoped, audited, and deprovisioned like any other principal, and it is distinct from every human principal and from other agent principals. Sponsor: The human principal (or accountable organizational principal) on whose behalf an agent session runs and to whom the session's authority traces. The sponsor is the originator of the delegation chain for the session. Session: A bounded episode of agent activity, initiated by or for a sponsor, with a defined start, end, and scope. The session is the authorization unit of this profile: credentials, effective permissions, and audit context are all scoped to it. Session Credential: A short-lived credential issued for a session, bound to the agent principal, the sponsor, and the session identifier. Session Scope: The set of tools, targets, and constraints authorized for a session, established when the session is created and modifiable thereafter only by narrowing, except through a new sponsor authorization. Policy Enforcement Point (PEP): The component in the request path of a tool call that evaluates the call against policy and applies the resulting disposition before the call reaches the tool. A gateway interposed between agents and tools is the common deployment form. Disposition: The outcome of evaluating a tool call: one of allow, deny, escalate, or redact (Section 4.5). Audit Receipt: The tamper-evident record emitted for each evaluated tool call, containing the fields specified in Section 4.8. Delegation Chain: The ordered sequence of principals -- sponsor, agent, and any sub-agents -- linked by authenticated delegation steps, under which a session's authority flows. Unscoreable Action: A tool call whose authorization evaluation cannot be completed: because the tool is unknown to the policy system, the arguments cannot be parsed or classified, the policy service is unavailable or in error, or required context is missing. Pandey Expires 21 May 2027 [Page 6] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 3. Architecture and Roles The profile involves five roles: o The Sponsor initiates a session (directly, or through a scheduler or another application acting under the sponsor's authority) and sets the ceiling on the session's authority. The sponsor is accountable for the session in the same way a resource owner is accountable for an OAuth grant. o The Agent acts under an agent principal. At session start the agent authenticates as its principal and obtains a session credential (Section 4.3). During the session it issues tool calls. o The Authorization Server issues session credentials, evaluates or supplies the policy inputs for authorization decisions, and provides revocation [RFC7009] and introspection [RFC7662] services. In cloud IAM terms this role is filled by the IAM control plane; in OAuth deployments, by the authorization server of [RFC6749]. o The Policy Enforcement Point sits in the path of every tool call. It authenticates the session credential, evaluates the call (Section 4.4), applies the disposition (Section 4.5), and emits the audit receipt (Section 4.8). Calls that do not pass the PEP do not reach the tool: a deployment in which the agent can reach tools by a path that bypasses the PEP does not implement this profile. o The Audit Service receives and retains audit receipts in a store that is outside the agent's write permissions and is tamper-evident (Section 4.8). +-------------------+ | Sponsor | +---------+---------+ | initiates session, | sets authority ceiling v +-------------------+ +--------------------+ | Agent Principal |------>| Authorization | | | sess. | Server | | | cred. | | +---------+---------+ +--------------------+ | tool call v +-------------------+ +--------------------+ | Policy |------>| Audit Service | | Enforcement Point | rcpt. | (tamper-evident, | | | | append-only) | +---------+---------+ +--------------------+ | allowed / redacted v +-------------------+ | Tool / Resource | | Server | Pandey Expires 21 May 2027 [Page 7] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 +-------------------+ Figure: Roles and flows in the session-scoped authorization model. An example of the PEP role is a gateway interposed between agents and the model and tool endpoints they call. The open-source Argus gateway, for example, applies a per-request policy decision (in its local vocabulary: allow, redact, block, or flag) to each request and records an audit entry for every decision [ARGUS]; implementations map local decision vocabularies onto the dispositions defined in Section 4.5. 4. Session-Scoped Authorization Model 4.1. Agents as First-Class Principals Each agent MUST be assigned its own agent principal. Distinct agent deployments MUST NOT share a principal, and an agent principal MUST NOT be shared with a human user or a non-agent workload. Agents MUST NOT access tools under shared service accounts. Agent principals SHOULD be provisioned from reviewed templates by an automated issuance process, not created by hand: the issuance rule is that every agent identity traces to a reviewed template and a recorded deployment. Template-based issuance is what makes the identity plane reviewable -- ad hoc identities cannot be audited. Per-agent principals are the precondition for everything that follows: intersection (Section 4.4) is meaningless if "the agent's permissions" are in fact the union held by a shared account, and attribution (Section 4.8) is impossible if several agents are indistinguishable in the log. 4.2. Sponsorship and Delegation Chains Every session MUST be bound to an authenticated sponsor. The binding is a delegation: the sponsor delegates a bounded subset of authority to the agent for the session. Delegation, in the sense of [RFC8693], preserves both identities -- the agent acts as itself, on behalf of the sponsor -- as distinct from impersonation, in which the delegatee becomes indistinguishable from the delegator. Implementations of this profile MUST use delegation semantics and MUST NOT use impersonation: an audit trail that cannot distinguish agent from sponsor has lost the property this document exists to provide. Where an agent delegates to a sub-agent, each delegation step: o MUST be authenticated, for example by an OAuth 2.0 Token Exchange [RFC8693] in which the resulting token's actor ("act") claim records the chain; o MUST NOT amplify authority: the delegatee's effective authority is bounded by the delegator's effective authority at the time of delegation, and can only narrow further down the chain (attenuation); Pandey Expires 21 May 2027 [Page 8] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 o MUST record the step such that the full chain, from sponsor to acting agent, is recoverable from audit receipts; o SHOULD be bounded in depth by policy; deployments SHOULD set an explicit maximum chain depth appropriate to their environment. The sponsor's authority is the ceiling for the entire chain: no combination of delegation steps can produce authority the sponsor does not hold (Section 4.4). Note that [RFC8693] treats nested actor information as history for audit rather than as enforced authority; enforcement of attenuation at each hop is a requirement of this profile, placed on the PEP and the authorization server, and token mechanisms that make attenuation itself verifiable are emerging [I-D.asor-wimse-agent-delegation-chain]. 4.3. Session-Scoped Credentials Credentials used to access tools during a session MUST be session credentials: issued for the session, bound to the agent principal, the sponsor, and the session identifier, and short-lived -- on the order of minutes to hours, never days. Session credentials: o MUST expire with or before the end of the session; o MUST NOT be persisted to disk, environment variables, or logs; o SHOULD be sender-constrained, using mutual TLS [RFC8705] or DPoP [RFC9449], so that a leaked credential cannot be replayed by another party; o are rotated by expiry: a new session obtains a new credential, and no long-lived secret exists to rotate operationally. Short-lived session credentials are standard practice in cloud IAM (temporary credentials issued per role assumption) applied to agents as an architectural rule. The security property is structural: there is no static secret to leak, and a compromised session cannot be replayed after it ends. 4.4. Per-Tool-Call Authorization and Permission Intersection When an agent acts on behalf of a sponsor, the agent's authority is exercised in addition to -- never instead of -- the sponsor's own authorization. The effective permission set for a session is: EffectivePermissions = AgentPermissions INTERSECT SponsorPermissions INTERSECT SessionScope The intersection rule is the load-bearing control of this profile. Its critical consequence concerns prompt injection: an agent's instructions arrive partly from untrusted sources (tool outputs, retrieved documents, web pages), and a malicious instruction embedded in any of them can redirect the agent to exercise permissions its Pandey Expires 21 May 2027 [Page 9] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 sponsor does not hold [GRESHAKE]. If the effective permission set is the intersection, the injected instruction cannot reach beyond what the sponsor could do directly; the ceiling is structural, not behavioral, and does not depend on the model resisting the instruction. Authorization MUST be evaluated for each tool call, at the PEP, before the call is executed -- not only when the session is established. Session-start evaluation is insufficient for agents because the actions to be authorized do not exist at session start: the agent selects each call's target and arguments at runtime. The evaluation of a tool call takes as inputs, at minimum: o the authenticated agent principal; o the sponsor and the delegation chain under which the session runs; o the session identifier and session scope; o the identity of the target tool (Section 6.2); o the arguments of the call, or a classification of them sufficient for policy evaluation. Policy systems MAY use contextual attributes (session tags, condition keys, time, task identifiers) as additional inputs. Policies and intersection semantics apply unchanged to calls made by sub-agents: the chain presented with the call determines the sponsor intersection, and a sub-agent's call is evaluated against its own (attenuated) position in the chain. 4.5. Dispositions and Fail-Closed Behavior Evaluation of a tool call produces exactly one disposition: allow: The call is executed as presented. deny: The call is not executed. The agent is informed that the call was denied, without disclosure of policy internals beyond what the deployment chooses to reveal. escalate: The call is suspended pending sponsor approval (Section 4.6). If approval is not granted, the disposition resolves to deny. redact: The call is executed with specified transformations applied -- for example, removal or masking of detected secrets or personal data from arguments or results -- where policy defines the call as permissible only in transformed form. Every evaluated call MUST be recorded in an audit receipt, regardless Pandey Expires 21 May 2027 [Page 10] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 of disposition (Section 4.8): recording is not itself a disposition, and "allow and log" is not a substitute for evaluation. An unscoreable action MUST be denied. Where policy defines an approval path for the action's class, an unscoreable action MAY be escalated instead; it MUST NOT be allowed, and it MUST NOT be passed through on the grounds that logging it suffices. Fail-closed is the default because the alternative -- executing actions the policy system could not evaluate -- converts every policy outage and every novel tool into an authorization bypass. 4.6. Human Approval for Irreversible Actions Policy SHOULD designate classes of actions as requiring sponsor approval before execution -- typically actions that are irreversible or high-impact in the deployment's terms (deletion of data, external communication, financial or physical-world effects). For such classes the PEP applies the escalate disposition. An approval, when granted, MUST be bound to the specific action: the tool target and a hash of the arguments as presented. An approval MUST NOT be a standing grant for a class of future calls, and it MUST expire promptly. This binds the human's decision to the action the human reviewed; an agent that obtains approval for one call cannot replay it for a different call with altered arguments. 4.7. Revocation A session's authority MUST be revocable at three speeds, in order of increasing finality: 1. Expiry. Session credentials expire on their own within minutes to hours. This is the routine mechanism and requires no action. 2. Explicit deny. An explicit deny attached to the agent principal (or to the session) takes immediate effect at the PEP and overrides any allow. This is the incident-response mechanism: one action, effective immediately. 3. Deprovisioning. The agent principal is deleted or its trust policy is narrowed, ending the identity rather than the session. Where OAuth tokens carry session context, revocation uses the mechanism of [RFC7009], and PEPs that validate tokens online use introspection [RFC7662] so that revocation is observed without waiting for expiry. Revocation of a sponsor's session MUST also terminate the authority of every delegation chain rooted in that session. Revocation is paired with a bounded blast-radius review: because every action of the session is attributable (Section 4.8), the operator can enumerate what the identity touched during its lifetime. That enumeration is what the audit trail exists to make possible. Pandey Expires 21 May 2027 [Page 11] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 4.8. Audit Receipts Every evaluated tool call MUST produce an audit receipt containing at least: o the agent principal identifier; o the sponsor identifier, and the delegation chain where the call was made under delegated authority; o the session credential identifier; o the tool target identifier; o a hash of the tool-call arguments; o the disposition, and the result status of the call if executed; o a timestamp. Receipts record the triple that shared service accounts erase: which agent performed which call, on whose behalf. Two properties matter beyond completeness: o Tamper evidence and separation. Receipts MUST be written to a store outside the agent's write permissions. The store SHOULD be append-only and tamper-evident (for example, by hash-chaining successive receipts). An agent that can edit its own audit trail has no audit trail. o Data minimization. Receipts carry a hash of the arguments rather than the arguments themselves, balancing forensic usefulness against the risk of the audit log becoming a sensitive-data store. Full payloads MAY be retained separately under stricter access controls where regulation or incident response requires it. 5. Protocol and Deployment Considerations 5.1. Token Contents and Identifiers The profile is transport- and format-agnostic, but the following assignments apply when OAuth 2.0 and JWT [RFC7519] carry the session context, consistent with [I-D.ietf-wimse-aims] and the JWT access token profile [RFC9068]: o The agent is identified as the OAuth client: the acting agent is carried in "client_id", and the agent's workload identity is established per [I-D.ietf-wimse-arch]. o The sponsor -- the party on whose behalf the agent acts -- is carried in the "sub" claim. o Delegation chains are carried in the "act" claim of [RFC8693], extended by token exchange at each hop. o The session is identified by a session identifier bound into the Pandey Expires 21 May 2027 [Page 12] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 session credential and echoed in audit receipts. Where the session identifier is expressed as a JWT claim, this document does not register a claim name; deployments use collision- resistant or private-use claim names as described in [RFC7519]. Registration of a standard session identifier claim is a possible subject of a future revision; it is deliberately not bundled into this document (Section 7). A deployment's identifiers MUST be stable for the life of a session and MUST appear identically in credentials, PEP decisions, and audit receipts; attribution fails if the join keys differ between the system that acts and the system that records. 5.2. Mapping to OAuth and WIMSE Mechanisms The model composes existing mechanisms as follows: o Identity. Workload identity and workload credentials per the WIMSE architecture [I-D.ietf-wimse-arch]; agent identity layered on OAuth and WIMSE per [I-D.ietf-wimse-aims]. o Delegation. Token exchange [RFC8693] with delegation (not impersonation) semantics, at session establishment and at each sub-agent hop. o Sender constraining. Session credentials bound to a key or certificate via DPoP [RFC9449] or mutual TLS [RFC8705], per the security best practice of [RFC9700]. o Revocation and status. Token revocation [RFC7009]; introspection [RFC7662] at PEPs that validate online. o Context propagation. Within a trust domain, transaction tokens [I-D.ietf-oauth-transaction-tokens] can carry the session context (sponsor, session, chain) alongside the call. o Baseline. The OAuth 2.0 security best current practice [RFC9700] applies throughout; the consolidated OAuth 2.1 framework [I-D.ietf-oauth-v2-1] does not change any requirement of this profile. 5.3. Cost Attribution Identity is the join key for cost governance; the point is brief but load-bearing, and is included because it falls out of the same mechanism. Attributing inference and tool spend to a task, an agent, and a tenant is meaningful only if "agent" and "tenant" in the accounting record are real identities rather than labels the application chose to write. When every tool call carries the agent principal and the sponsor (Section 4.8), spend attribution becomes a query over the audit trail instead of a reconstruction project. Spending caps therefore attach to the principal as policy, on the same plane as permission policy, and inherit its properties: Pandey Expires 21 May 2027 [Page 13] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 evaluated per invocation, fail-closed, and not circumventable by the agent. The intersection principle extends directly: a session's effective spend authority is the minimum of the agent principal's budget and the sponsor's (or tenant's) budget -- an agent serving a cost-constrained sponsor cannot outspend the sponsor's limit whatever its own principal allows. The composition of identity with cost governance is developed in [PANDEY-IAM] and in the reference architecture of [PANDEY-REFARCH]. 5.4. Deployment Experience and Open Issues The requirements above are stated at the strength the underlying experience supports [PANDEY-IAM]. In the production deployments described there, per-agent principals, the intersection rule, and attributable audit trails are deployed practice; session-scoped credentials are standard cloud IAM practice applied to agents. Three items are stated as requirements of the profile but are identified in [PANDEY-IAM] as less mature in deployment, and are recorded here as open issues for implementers and for future revisions of this document: o Policy-level approval enforcement. Approval workflows for irreversible actions exist in deployments as application-layer conventions. Full policy-level enforcement of the bound approval tokens required by Section 4.6 is a recommendation in the source deployments, not yet their operating practice. o Emergency revocation under incident conditions. The explicit-deny mechanism of Section 4.7 is defined and available in the deployments described, but single-action emergency revocation is reported as untested under live incident conditions. o Delegation chains at depth. The intersection rule composes transitively in principle, and Section 4.2 states the resulting requirements. The policy mechanics of chained session context -- agents invoking agents invoking agents -- have not, to the author's knowledge, been demonstrated in production at depth, by the source deployments or others. This is the most important open problem in this design space; mechanism work such as [I-D.asor-wimse-agent-delegation-chain] is directed at it. Deployments that evaluate the intersection only at session establishment, rather than per tool call, do not conform to this profile (Section 4.4); the per-call evaluation requirement imposes latency and policy-availability costs that implementers should budget for explicitly (see also Section 6.7). 6. Security Considerations This section follows the guidance of [RFC3552]. The threat model assumes the IAM control plane, the authorization server, the PEP, and the audit store are trusted, and that content the agent reads -- tool outputs, retrieved documents, web pages, files -- is not. Five agent-specific threats organize the analysis; each is stated with the Pandey Expires 21 May 2027 [Page 14] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 control this profile applies and the residual risk that remains. 6.1. Prompt Injection as Privilege Escalation An agent's instructions arrive partly from untrusted sources. A malicious instruction embedded in a tool result or a retrieved document can redirect the agent to exercise permissions its sponsor does not hold [GRESHAKE]; the delivery mechanism is natural language, but the effect is privilege escalation. The intersection rule (Section 4.4) bounds the effect structurally: injected instructions cannot reach beyond the sponsor's own authority, because the agent never holds more. Residual risk: intersection does not prevent an agent from misusing authority the sponsor legitimately holds. A poisoned document that induces the agent to misuse its sponsor's own permissions -- deleting the sponsor's files with the sponsor's authority -- is a data-plane attack that authorization cannot, by itself, answer. Input validation at the tool boundary, argument classification at the PEP, and the escalate disposition for irreversible classes (Section 4.6) are the mitigating controls; eliminating the risk class is out of scope for this document. 6.2. Tool Squatting and Tool Substitution An agent selects tools by name and description from a registry or discovery mechanism. An attacker who can register a tool under a confusable name, or alter a tool's description to misrepresent its behavior, can redirect calls -- with the session's full authority -- to an attacker-controlled endpoint. PEPs and registries MUST authenticate tool identity: the tool target recorded in policy and in audit receipts MUST be the authenticated identity of the tool, not a display name or an unauthenticated description. Tool metadata used in evaluation SHOULD be integrity-protected, and changes to a tool's registered behavior SHOULD be treated as a policy-relevant event requiring re-review. 6.3. Confused Deputy The classic confused-deputy pattern applies with a language-driven twist: the deputy (the agent) is confused not by forged references but by natural-language instructions from untrusted sources, and it holds ambient authority if run under a shared account. Per-agent principals remove the ambient authority; intersection removes the escalation path; session scope removes the cross-task reach. A PEP MUST NOT treat the agent's request as self-authenticating evidence of sponsor intent: authority derives from the session credential and the delegation chain, never from assertions in the call's content (including claims in arguments that the sponsor approved the action). 6.4. Token Theft, Replay, and Long-Lived Credentials Bearer session credentials, if leaked, are usable by any holder until expiry. Short lifetimes (Section 4.3) bound the window; sender Pandey Expires 21 May 2027 [Page 15] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 constraining [RFC8705] [RFC9449] binds the credential to a key the attacker does not hold and is the effective control; [RFC9700] recommends sender constraining for exactly this reason. Long-lived agent credentials MUST NOT be issued under this profile: the credential's lifetime is the rotation policy, and a credential that outlives its session is a standing compromise waiting for a log file, an environment dump, or a misconfigured secret store to expose it. Credentials MUST NOT be written to logs -- including, specifically, to audit receipts, which carry credential identifiers, never credential material. 6.5. Delegation Chain Amplification and Splicing Delegation introduces its own attacks. Amplification -- a hop that grants more than it holds -- is prevented by the attenuation requirement of Section 4.2, enforced by the authorization server at exchange time and by the PEP at call time. Splicing -- combining chain segments from different sessions to assemble authority no single session granted -- is prevented by binding each chain to its session identifier and sponsor, and by PEP verification that the chain presented with a call belongs to the session the credential was issued for. Depth limits bound the audit and evaluation cost of pathological chains. Operators should note the asymmetry the profile relies on: [RFC8693] actor information is, by itself, audit history; attenuation must be enforced, not merely recorded. 6.6. Audit Record Integrity and Privacy The audit trail is a security control only if it survives the incident it is meant to explain. Receipts are written outside the agent's permissions and SHOULD be hash-chained or otherwise made tamper-evident (Section 4.8); compromise of the agent must not imply compromise of the record. Denial of audit -- an agent or attacker flooding or suppressing receipt generation -- is mitigated by the PEP's position in the request path: no receipt, no call. The argument-hash design limits the privacy exposure of the trail itself; deployments retaining full arguments are creating a sensitive-data store and MUST protect it accordingly, including access control and retention limits. Timestamps and identifiers in receipts can reveal work patterns of sponsors; access to receipts SHOULD be restricted to those with an operational or audit need. 6.7. Fail-Closed Denial of Service Fail-closed behavior (Section 4.5) creates an availability lever: an attacker who can make actions unscoreable -- by feeding the agent malformed inputs, novel tools, or pathological arguments -- can block work. This is the correct failure direction for an authorization system (the alternative converts policy failure into silent privilege), but deployments SHOULD monitor denial rates, alert on evaluation-service failure as a security event rather than an outage to be routed around, and MUST NOT respond to evaluation outages by failing open. Where continuity matters, the escalate path -- human approval for a suspended call -- is the designed pressure valve, not Pandey Expires 21 May 2027 [Page 16] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 an allow-by-default fallback. 6.8. Sponsor Compromise and Sponsor Authentication Every guarantee in this profile is bounded by the sponsor's authentication: an attacker holding the sponsor's credentials holds the ceiling. Sponsor authentication SHOULD be phishing-resistant, consistent with [RFC9700]. Session establishment is the moment of maximum leverage and SHOULD be logged and reviewable with the same rigor as the tool calls it enables. Accountability cuts both ways: because receipts bind actions to a sponsor, a sponsor is answerable for sessions they originate, which is the intended property, and deployments should make that visible to sponsors at session start rather than discovering it during an incident. 7. IANA Considerations This document has no IANA actions. It registers no parameters, claims, media types, or other registry entries. The session identifier discussed in Section 5.1 is carried using existing or private-use mechanisms, as described there. 8. References 8.1. Normative References [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, . [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, . [RFC6749] Hardt, D., Ed., "The OAuth 2.0 Authorization Framework", RFC 6749, DOI 10.17487/RFC6749, October 2012, . [RFC7009] Lodderstedt, T., Ed., Dronia, S., and M. Scurtescu, "OAuth 2.0 Token Revocation", RFC 7009, DOI 10.17487/RFC7009, August 2013, . [RFC7519] Jones, M., Bradley, J., and N. Sakimura, "JSON Web Token (JWT)", RFC 7519, DOI 10.17487/RFC7519, May 2015, . [RFC7662] Richer, J., Ed., "OAuth 2.0 Token Introspection", RFC 7662, DOI 10.17487/RFC7662, October 2015, . [RFC8693] Jones, M., Nadalin, A., Campbell, B., Ed., Bradley, J., and C. Mortimore, "OAuth 2.0 Token Exchange", RFC 8693, DOI 10.17487/RFC8693, January 2020, . Pandey Expires 21 May 2027 [Page 17] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 [RFC8705] Campbell, B., Bradley, J., Sakimura, N., and T. Lodderstedt, "OAuth 2.0 Mutual-TLS Client Authentication and Certificate-Bound Access Tokens", RFC 8705, DOI 10.17487/RFC8705, February 2020, . [RFC9068] Bertocci, V., "JSON Web Token (JWT) Profile for OAuth 2.0 Access Tokens", RFC 9068, DOI 10.17487/RFC9068, October 2021, . [RFC9449] Fett, D., Campbell, B., Bradley, J., Lodderstedt, T., Jones, M., and D. Waite, "OAuth 2.0 Demonstrating Proof of Possession (DPoP)", RFC 9449, DOI 10.17487/RFC9449, September 2023, . [RFC9700] Lodderstedt, T., Bradley, J., Labunets, A., and D. Fett, "Best Current Practice for OAuth 2.0 Security", BCP 240, RFC 9700, DOI 10.17487/RFC9700, January 2025, . 8.2. Informative References [I-D.ietf-wimse-arch] Salowey, J., Rosomakho, Y., and H. Tschofenig, "Workload Identity in a Multi System Environment (WIMSE) Architecture", Work in Progress, Internet-Draft, draft-ietf-wimse-arch-08, 6 July 2026, . [I-D.ietf-wimse-aims] Kasselman, P., Lombardo, J., Rosomakho, Y., Campbell, B., Steele, N., and A. Parecki, "AI Identity Management System", Work in Progress, Internet-Draft, draft-ietf-wimse-aims-01, 2 November 2026, . [I-D.ni-wimse-ai-agent-identity] Ni, Y., Liu, C. P., Richardson, M., and B. Wang, "WIMSE Applicability for AI Agents", Work in Progress, Internet-Draft, draft-ni-wimse-ai-agent-identity-04, 27 October 2026, . [I-D.ietf-oauth-transaction-tokens] Tulshibagwale, A., Fletcher, G., and P. Kasselman, "Transaction Tokens", Work in Progress, Internet-Draft, draft-ietf-oauth-transaction-tokens-11, 30 July 2026, . [I-D.ietf-oauth-v2-1] Hardt, D., Parecki, A., Ed., and T. Lodderstedt, "The OAuth 2.1 Authorization Framework", Work in Progress, Internet-Draft, draft-ietf-oauth-v2-1-17, 31 October 2026, . [I-D.asor-wimse-agent-delegation-chain] Asor, R., "Verifiable Attenuated Delegation for AI Agent Chains", Work in Progress, Internet-Draft, draft-asor-wimse-agent-delegation-chain-03, 27 October 2026, Pandey Expires 21 May 2027 [Page 18] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 . [RFC3552] Rescorla, E. and B. Korver, "Guidelines for Writing RFC Text on Security Considerations", BCP 72, RFC 3552, DOI 10.17487/RFC3552, July 2003, . [GRESHAKE] Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., and M. Fritz, "Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection", arXiv:2302.12173, February 2023, . [SAGA] Syros, G., Suri, A., Ginesin, J., Nita-Rotaru, C., and A. Oprea, "SAGA: A Security Architecture for Governing AI Agentic Systems", arXiv:2504.21034, April 2025, . [NIST-NCCOE-AI-ID] Booth, H., Fisher, W., Galluzzo, R., and J. Roberts, "Accelerating the Adoption of Software and AI Agent Identity and Authorization", NIST National Cybersecurity Center of Excellence Concept Paper (Draft), February 2026. [NIST-AI-600-1] National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile", NIST AI 600-1, July 2024, DOI 10.6028/NIST.AI.600-1, . [AGENTCORE] Amazon Web Services, "Amazon Bedrock AgentCore Identity", in "Amazon Bedrock AgentCore Developer Guide", 2026, . [ENTRA-AGENT-ID] Microsoft, "Overview of agent identities in Microsoft Entra", Microsoft Learn, 2026, . [OWASP-LLM-TOP10] OWASP GenAI Security Project, "OWASP Top 10 for LLM Applications", 2026, . [PANDEY-IAM] Pandey, K., "Your AI Agent Is Now an IAM Principal: Identity Architecture for Autonomous Agents", Zenodo, October 2026, DOI 10.5281/zenodo.23420912, . [PANDEY-REFARCH] Pandey, K., "From Pilots to Production: A Reference Architecture for Cost-Governed AI Agents on AWS", Zenodo, October 2026, DOI 10.5281/zenodo.23178788, . [ARGUS] Pandey, K., "Argus: an OpenAI-compatible policy gateway for LLM requests", software, Version 0.1.2, October 2026, DOI Pandey Expires 21 May 2027 [Page 19] Internet-Draft Session-Scoped Auth. for AI Agents November 2026 10.5281/zenodo.23737764, . Author's Address Karmendra Pandey TEKsystems, AI & ML Practice Email: karmendra8386@gmail.com URI: https://orcid.org/0009-0000-3461-9889 Pandey Expires 21 May 2027 [Page 20]