<?xml version="1.0" encoding="utf-8"?>
<rfc xmlns:xi="http://www.w3.org/2001/XInclude"
     category="info"
     docName="draft-correa-agent-capability-binding-00"
     ipr="trust200902"
     submissionType="independent"
     version="3">
  <front>
    <title abbrev="Agent Capability Enforcement">Enforcement Profile for Agent Capabilities: Argument Constraints, Taint Propagation, and Consent Binding</title>
    <seriesInfo name="Internet-Draft" value="draft-correa-agent-capability-binding-00"/>
    <author fullname="Cleiton Augusto Correa Bezerra" initials="C.A." surname="Correa Bezerra">
      <address>
        <email>augusto.cleiton@gmail.com</email>
      </address>
    </author>
    <date year="2026" month="November"/>
    <area>Security</area>
    <workgroup>Independent Submission</workgroup>
    <keyword>AI agents</keyword>
    <keyword>authorization</keyword>
    <keyword>prompt injection</keyword>
    <keyword>capabilities</keyword>
    <keyword>receipts</keyword>
    <abstract>
      <t>AI agents increasingly call tools and hand work to other agents while holding broad, inherited authority. Most of the resulting incidents share one shape: an agent, often steered by injected content, performs an action its principal never authorized, and no party can prove the limit afterward. This document defines an interoperable enforcement profile that composes two existing building blocks, a delegable signed capability (the Attenuating Agent Tokens profile) and a replayable action receipt, and adds three enforcement rules that catch the common failures: argument-level constraints (including an egress host allowlist), taint propagation across a delegation chain, and a consent binding between a displayed decision and the action dispatched. Enforcement is specified at the tool-call boundary, outside the agent, so that a prompt cannot edit the limit and a verdict means the same thing across agent frameworks.</t>
    </abstract>
  </front>
  <middle>
    <section>
      <name>Introduction</name>
      <t>An AI agent is typically granted the full authority of the human or service that runs it, and that authority flows unchanged to every tool it calls and every subagent it delegates to. Reported 2026 incidents, a single email that takes over a cloud agent, a single prompt that reaches over-permissioned cloud credentials, assistants that exfiltrate data through indirect prompt injection, differ in surface but share a root cause: nothing enforces the principal's limit at the moment of the action, and the agent itself is the thing being fooled.</t>
      <t>Frameworks are responding with in-process permission gates, for example a capability list checked before tool invocation. This raises the bar but leaves three gaps: an in-process check can be routed around by another tool or a compromised step; each framework rebuilds its own, so a verdict does not travel when a crew hands off to a non-native agent; and a tool-level allowlist does not constrain the arguments where the real risk usually lives.</t>
      <t>This document does not define a new token or receipt format. It profiles two existing drafts, <xref target="AAT"/> for a signed, delegable (attenuating) capability and <xref target="RECEIPTS"/> for a replayable action receipt, and specifies three enforcement rules a conforming verifier applies at the tool-call boundary, outside the agent:</t>
      <ol>
        <li>Argument constraints, including an egress host allowlist (<xref target="argument-constraints"/>).</li>
        <li>Taint propagation across the delegation chain (<xref target="taint"/>).</li>
        <li>Consent binding between the displayed decision and the dispatched action (<xref target="consent"/>).</li>
      </ol>
      <t>The goal is interoperability: the same capability and the same receipt mean the same thing whether enforced in one framework, another, or a protocol-level tool host, so governance survives delegation across framework boundaries.</t>
    </section>
    <section>
      <name>Conventions and Definitions</name>
      <t>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 <xref target="RFC2119"/> <xref target="RFC8174"/> when, and only when, they appear in all capitals, as shown here.</t>
      <dl>
        <dt>Principal:</dt>
        <dd>the human or service whose authority a capability conveys.</dd>
        <dt>Verifier:</dt>
        <dd>the party that admits or refuses a tool call at the call boundary. The verifier is outside the agent and MUST NOT rely on state the agent can edit.</dd>
        <dt>Capability:</dt>
        <dd>a signed, attenuating token per <xref target="AAT"/> naming the tools, arguments, and limits an agent may use, derivable only to a subset by its holder.</dd>
        <dt>Receipt:</dt>
        <dd>a record per <xref target="RECEIPTS"/> of a tool call and its verdict, replayable by an auditor.</dd>
      </dl>
    </section>
    <section>
      <name>Enforcement Model</name>
      <t>A conforming verifier sits at the tool-call boundary and, for each call, evaluates the presented capability chain and the call's arguments. It admits the call only if all of the following hold, and it writes a receipt either way:</t>
      <ul>
        <li>the capability chain verifies and grants the tool (per <xref target="AAT"/>);</li>
        <li>every constrained argument satisfies its constraint (<xref target="argument-constraints"/>);</li>
        <li>no holder in the chain is tainted when the tool acts on the world (<xref target="taint"/>);</li>
        <li>any consent binding present matches the dispatched action (<xref target="consent"/>).</li>
      </ul>
      <t>The verifier MUST be reachable only through a path the agent's process does not control; a verdict computed inside the agent's process does not satisfy this profile.</t>
    </section>
    <section anchor="argument-constraints">
      <name>Argument Constraints</name>
      <t>A capability MAY constrain individual argument values, not only the tool name. The dangerous property of a call is usually carried by an argument, the host a fetch reaches, the recipient or amount of a payment, so a tool-level allowlist alone is insufficient.</t>
      <t>A verifier MUST support at least exact-match and set-membership constraints. A verifier that admits tools taking a URL, a webhook, or a redirect argument MUST support a host allowlist constraint ("host_in"): the argument value MUST parse as a URL whose host, compared case-insensitively and with any port ignored, is a member of the constraint's host set. The host MUST be determined by a URL parser that resolves userinfo, so that "https://allowed.example@attacker.example/" has host "attacker.example" and is refused. An empty host set admits nothing. A derived capability MAY only narrow a host set to a subset.</t>
      <t>This rule refuses the exfiltration-by-fetch step common to 2026 injection incidents even when the tool itself is granted: the injected agent may call the fetch tool, but only to a host the principal signed.</t>
    </section>
    <section anchor="taint">
      <name>Taint Propagation</name>
      <t>A capability MAY mark a tool as bringing in outside content ("tainting") or as acting on the world ("clean-only").</t>
      <t>When a tainting tool returns, every holder in the call's delegation chain is tainted. While any holder in a call's chain is tainted, the verifier MUST refuse a clean-only tool, regardless of which tool it is, until a principal clears the taint. Clearing MUST be an explicit human decision recorded as a receipt; delegating to a fresh subagent MUST NOT clear it.</t>
      <t>This rule addresses indirect prompt injection at the action layer: an agent that read untrusted content cannot send, pay, or fetch until a human reviews, so a successful injection buys no action. It does not prevent the injection; it prevents the action the injection needs.</t>
    </section>
    <section anchor="consent">
      <name>Consent Binding</name>
      <t>When a human approves an action from a rendered surface, the approval MAY carry a consent binding: a digest over the resolved action and the displayed value of each field that the action depends on, computed over the JCS form <xref target="RFC8785"/> of those fields.</t>
      <t>Before dispatching the approved action, the verifier MUST recompute the digest over the action it is about to perform and the values that were displayed, and MUST refuse on any mismatch. A digest over the canonical action alone is insufficient: it proves the action did not change, not that the surface showed that action. Binding the displayed value to the action closes the "show X, settle Y" gap, where a conformant pipeline displays one amount and dispatches another.</t>
      <t>Volatile inputs (for example a time-boxed exchange rate) SHOULD be carried as a separately bound quote with its own digest and validity window, referenced by the action, so that a re-quote is a new quote rather than a mismatch, while the value the human saw remains bound.</t>
    </section>
    <section>
      <name>Receipts and Reason Codes</name>
      <t>Every admitted or refused call MUST produce a receipt per <xref target="RECEIPTS"/>. A refusal receipt MUST carry a reason code. A verifier MUST distinguish a denial (the capability does not grant the call) from an escalation (an otherwise-granted irreversible action parked for a human gate); folding an escalation into a denial hides why the call paused and encourages rubber-stamping. These are different verdicts and MUST use different reason codes.</t>
    </section>
    <section>
      <name>Conformance Vectors</name>
      <t>A conforming verifier MUST refuse each of the following and admit the accept case, producing a receipt in every case:</t>
      <ol>
        <li>A payment above the capability's cumulative cap (over-cap).</li>
        <li>A payment to a recipient not in the capability (wrong payee).</li>
        <li>A payment above the action's amount with no signed overage policy (overpayment).</li>
        <li>A settlement of the exact amount whose timestamp falls outside the bound quote's window (expired rate lock).</li>
        <li>A fetch to a host not in the "host_in" set, including the "allowed.example@attacker.example" userinfo form.</li>
        <li>A clean-only call while a holder in the chain is tainted.</li>
        <li>A dispatched action whose consent digest does not match the displayed values (show X, settle Y).</li>
      </ol>
      <t>Accept: a call within cap, to an allowed payee or host, with no taint and a matching consent binding.</t>
    </section>
    <section>
      <name>Security Considerations</name>
      <t>This profile constrains the integrity of the action about to be performed, not the authenticity of what a human's eyes saw. A renderer that lies before approval is out of scope and requires a surface the verifier owns; clear-signing mechanisms address the third-party-surface case. The verifier is trusted and MUST run outside the agent's process, under a different operating-system principal where the agent has shell access, or an agent that can run commands as the verifier can read its key and lift its own limits. Receipts carry digests and pseudonyms, not the data a call touched.</t>
    </section>
    <section>
      <name>IANA Considerations</name>
      <t>This document has no IANA actions. Argument constraint type names used here ("host_in") are expected to be registered by the referenced capability profile.</t>
    </section>
  </middle>
  <back>
    <references>
      <name>Normative References</name>
      <reference anchor="RFC2119" target="https://www.rfc-editor.org/info/rfc2119">
        <front>
          <title>Key words for use in RFCs to Indicate Requirement Levels</title>
          <author initials="S." surname="Bradner" fullname="S. Bradner"/>
          <date year="1997" month="March"/>
        </front>
        <seriesInfo name="BCP" value="14"/>
        <seriesInfo name="RFC" value="2119"/>
      </reference>
      <reference anchor="RFC8174" target="https://www.rfc-editor.org/info/rfc8174">
        <front>
          <title>Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words</title>
          <author initials="B." surname="Leiba" fullname="B. Leiba"/>
          <date year="2017" month="May"/>
        </front>
        <seriesInfo name="BCP" value="14"/>
        <seriesInfo name="RFC" value="8174"/>
      </reference>
      <reference anchor="RFC8785" target="https://www.rfc-editor.org/info/rfc8785">
        <front>
          <title>JSON Canonicalization Scheme (JCS)</title>
          <author initials="A." surname="Rundgren" fullname="A. Rundgren"/>
          <author initials="B." surname="Jordan" fullname="B. Jordan"/>
          <author initials="S." surname="Erdtman" fullname="S. Erdtman"/>
          <date year="2020" month="June"/>
        </front>
        <seriesInfo name="RFC" value="8785"/>
      </reference>
    </references>
    <references>
      <name>Informative References</name>
      <reference anchor="AAT" target="https://datatracker.ietf.org/doc/draft-niyikiza-oauth-attenuating-agent-tokens/">
        <front>
          <title>Attenuating Agent Tokens (AAT)</title>
          <author>
            <organization>IETF</organization>
          </author>
          <date year="2026"/>
        </front>
      </reference>
      <reference anchor="RECEIPTS" target="https://datatracker.ietf.org/doc/draft-sahu-agent-action-receipts/">
        <front>
          <title>Agent Action Receipts</title>
          <author>
            <organization>IETF</organization>
          </author>
          <date year="2026"/>
        </front>
      </reference>
    </references>
    <section numbered="false">
      <name>Acknowledgments</name>
      <t>This profile composes and depends on the Attenuating Agent Tokens and Agent Action Receipts drafts; the consent-binding rule was sharpened in public review with payments and agent-UI practitioners.</t>
    </section>
  </back>
</rfc>
