Independent Submission W. Han
Internet-Draft Individual
Intended status: Informational 17 November 2026
Expires: 21 May 2027
AI Manifest: Site-Published Friction-Recovery Descriptors for AI Agents
draft-han-ai-manifest-04
Abstract
This document specifies the AI Manifest, a JSON format with which a
website operator publishes an AI Friction-Recovery Manifest (AFRM): a
description of the user-interface (UI) patterns on its own origin
that predictably stall autonomous AI agents, together with
constrained recovery steps and site-specific input hints. An agent
that drives a browser on behalf of a user retrieves the manifest from
a well-known URI and consults it when it encounters friction, instead
of rediscovering the same site-specific behavior by trial and error.
The manifest is untrusted input, interpreted by the agent's runtime
and never shown to its language model. Its operative content is
drawn from a closed vocabulary that cannot enter text, upload files,
or reach another origin. Every action derived from a manifest passes
through a policy enforcement point that is not a language model, and
steps conventionally bound to a decision always require the user's
confirmation; an agent's own policy can only deny them. Security
controls such as access denials, rate limits, CAPTCHAs,
authentication challenges, and confirmation dialogs are stop
conditions, never traps to be recovered from. Optional publisher
signatures and agent-configured reputation services provide
authenticity but never imply safety and never relax these
requirements.
Preliminary measurements, including an adversarial evaluation of a
reference implementation, are reported in an appendix.
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/.
Han Expires 21 May 2027 [Page 1]
Internet-Draft AI Manifest November 2026
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.
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.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1. Problem Statement . . . . . . . . . . . . . . . . . . . . 3
1.2. Applicability . . . . . . . . . . . . . . . . . . . . . . 4
1.3. Relationship to Other Work . . . . . . . . . . . . . . . 4
1.4. Design Principles . . . . . . . . . . . . . . . . . . . . 5
2. Conventions and Terminology . . . . . . . . . . . . . . . . . 6
3. Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.1. Well-Known URI . . . . . . . . . . . . . . . . . . . . . 7
3.2. Link Relation . . . . . . . . . . . . . . . . . . . . . . 8
3.3. Removed Discovery Methods . . . . . . . . . . . . . . . . 8
3.4. Agent-Local Manifests . . . . . . . . . . . . . . . . . . 9
3.5. Retrieval and Caching . . . . . . . . . . . . . . . . . . 9
4. Manifest Format . . . . . . . . . . . . . . . . . . . . . . . 10
4.1. Top-Level Members . . . . . . . . . . . . . . . . . . . . 10
4.2. Locator Object . . . . . . . . . . . . . . . . . . . . . 11
4.3. Trap Descriptor . . . . . . . . . . . . . . . . . . . . . 12
4.4. Step Object . . . . . . . . . . . . . . . . . . . . . . . 12
4.5. Field Hint Descriptor . . . . . . . . . . . . . . . . . . 13
4.6. Shortcut Descriptor . . . . . . . . . . . . . . . . . . . 14
4.7. Manifest Strings and the Language Model . . . . . . . . . 15
4.8. Vocabularies . . . . . . . . . . . . . . . . . . . . . . 15
4.8.1. Actions . . . . . . . . . . . . . . . . . . . . . . . 15
4.8.2. Trap Categories . . . . . . . . . . . . . . . . . . . 16
4.8.3. Rendering Models . . . . . . . . . . . . . . . . . . 16
5. Agent Processing Requirements . . . . . . . . . . . . . . . . 16
5.1. Consulting the Manifest . . . . . . . . . . . . . . . . . 16
5.2. Policy Enforcement Point . . . . . . . . . . . . . . . . 18
5.3. Sensitive Actions . . . . . . . . . . . . . . . . . . . . 20
Han Expires 21 May 2027 [Page 2]
Internet-Draft AI Manifest November 2026
5.4. Security Controls Are Not Friction . . . . . . . . . . . 21
5.5. Audit Record . . . . . . . . . . . . . . . . . . . . . . 22
6. Authenticity and Reputation . . . . . . . . . . . . . . . . . 22
6.1. Publisher Signatures . . . . . . . . . . . . . . . . . . 22
6.2. Reputation Services . . . . . . . . . . . . . . . . . . . 22
7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 23
7.1. Well-Known URI Registration . . . . . . . . . . . . . . . 23
7.2. Link Relation Type Registration . . . . . . . . . . . . . 23
7.3. No New Registries . . . . . . . . . . . . . . . . . . . . 24
8. Security Considerations . . . . . . . . . . . . . . . . . . . 24
8.1. Threat Model . . . . . . . . . . . . . . . . . . . . . . 24
8.2. Prompt Injection . . . . . . . . . . . . . . . . . . . . 24
8.3. Exfiltration . . . . . . . . . . . . . . . . . . . . . . 25
8.4. Locator Substitution . . . . . . . . . . . . . . . . . . 25
8.5. Publisher-Principal Conflict . . . . . . . . . . . . . . 26
8.6. Circumvention of Security Controls . . . . . . . . . . . 26
8.7. Local Configuration Supply Chain . . . . . . . . . . . . 27
8.8. Information Disclosure, Reconnaissance, and Cloaking . . 27
8.9. Distinguishing Conforming Agents . . . . . . . . . . . . 28
8.10. Resource Exhaustion and Transport . . . . . . . . . . . . 28
9. Privacy Considerations . . . . . . . . . . . . . . . . . . . 28
10. Implementation Status . . . . . . . . . . . . . . . . . . . . 28
11. Intellectual Property Rights Disclosure . . . . . . . . . . . 29
12. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 29
13. References . . . . . . . . . . . . . . . . . . . . . . . . . 30
13.1. Normative References . . . . . . . . . . . . . . . . . . 30
13.2. Informative References . . . . . . . . . . . . . . . . . 32
Appendix A. Example Manifest . . . . . . . . . . . . . . . . . . 33
Appendix B. Preliminary Measurements . . . . . . . . . . . . . . 35
Appendix C. Document History . . . . . . . . . . . . . . . . . . 39
Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 41
1. Introduction
1.1. Problem Statement
AI agents increasingly operate websites on behalf of users by driving
a browser through automation interfaces such as WebDriver, the Chrome
DevTools Protocol, or tools exposed to a language model through the
Model Context Protocol [MCP]. Such agents observe a page through its
Document Object Model (DOM), its accessibility tree, or screenshots,
and infer how to operate it.
Han Expires 21 May 2027 [Page 3]
Internet-Draft AI Manifest November 2026
On many sites, and especially on enterprise transactional
applications, that inference repeatedly fails at the same places: a
cell edit that must be committed with a key gesture before the row
unlocks, a list that renders only the visible window, a field whose
valid values depend on configuration that is not visible on the page.
Each agent rediscovers these patterns by trial and error, at a cost
in actions, tokens, and failed tasks.
The AI Manifest lets the operator of an origin describe these
patterns once, in a form that an agent can consult and that an
agent's safety machinery can check mechanically.
1.2. Applicability
A manifest is appropriate only for friction that meets at least one
of the following conditions:
1. The UI cannot practically be changed by the party able to
describe it, for example a deployed vendor application whose
source is not accessible to the deploying organization.
2. The friction remains valid for human users and exists for a
reason, for example a commit gesture that protects data
integrity, so that removing it would be wrong even though it
stalls agents.
3. Operating the UI correctly requires site-specific knowledge that
is not derivable from public or training data, for example which
code values are registered for a given organizational unit.
Where a site can expose a structured interface to agents -- an HTTP
API, an MCP server [MCP], page-registered tools [WEBMCP], or an agent
card [A2A] -- that interface is preferable and agents SHOULD use it
instead of UI-level recovery. The AI Manifest addresses the residual
UI that such interfaces do not cover. Friction that consists of a
security control is never in scope (Section 5.4).
In the first case the party that publishes the manifest is typically
the deploying organization rather than the application vendor. Such
an organization usually cannot change the application, but does
control the web server or reverse proxy in front of it, and can serve
a static file under /.well-known/ on that origin.
1.3. Relationship to Other Work
Han Expires 21 May 2027 [Page 4]
Internet-Draft AI Manifest November 2026
* robots.txt [RFC9309] and the vocabulary work of the IETF AIPREF
working group [AIPREF] express whether and how automated clients
may use content. The AI Manifest does not grant or express
permission; an agent that consults it remains subject to those
mechanisms.
* llms.txt [LLMSTXT] provides free-text documentation for language
models. The AI Manifest is deliberately not free text: its
operative content is a closed vocabulary that can be validated
without a language model.
* Page-registered tools [WEBMCP] let a site execute functions on
behalf of an agent. The AI Manifest never executes anything; it
describes how the existing UI behaves.
* Agent authentication work such as the IETF Web Bot Auth working
group [WEBBOTAUTH] lets a site identify agents. It is
complementary (Section 8.8).
1.4. Design Principles
The security model of this revision rests on six principles. Every
normative requirement below derives from one of them.
P1. Untrusted by construction. A manifest is authored by the site,
whose interests may differ from the user's. Its delivery channel,
signature, or reputation status never makes its content trusted.
P2. Data, not instructions. Operative content is expressed in a
closed vocabulary of actions and categories, interpreted by the
agent's runtime. No manifest string reaches the language model.
P3. Origin-bound. A manifest describes only the origin that serves
it, and actions derived from it can act only within that origin.
P4. Least authority. A manifest can inform or narrow what an agent
does; it can never make an agent do something it would not
otherwise be permitted to do, and never removes a user
confirmation.
P5. Security controls are not friction. Mechanisms that deny,
challenge, throttle, or ask for human confirmation are stop
conditions, not traps to be escaped.
P6. Mediated outside the model. Every manifest-derived action
passes, before it has any effect, through a policy enforcement
point whose decision a language model cannot influence, and is
recorded.
Han Expires 21 May 2027 [Page 5]
Internet-Draft AI Manifest November 2026
P5 and P6 draw on the published design and defensive analysis of an
open-source autonomous penetration-testing system [ARTEX]. Its
defense guide identifies persistence with bypass variations after
401, 403, or 429 responses as a behavioral indicator of autonomous
attack, and its tool layer classifies such responses, including WAF
blocks, rate limits, and CAPTCHAs, as "blocked" so that its planner
changes strategy rather than stopping. A friction-recovery mechanism
that did not exclude security controls would teach benign agents the
same behavior. The same system passes every tool call through a hook
that evaluates priority-ordered rules with allow, deny, and ask
outcomes, optionally falls back to a language-model judge, and
records an audit log. Section 5.2 adopts that structure for
manifest-derived actions, but excludes a language model from the
decision.
2. Conventions and Terminology
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.
AI Manifest; AI Friction-Recovery Manifest (AFRM)
A JSON object [RFC8259] conforming to Section 4. The two terms
are used interchangeably.
Publisher
The origin [RFC6454] that serves a manifest, and the operator
responsible for it.
Agent
Software, typically driven by a large language model (LLM), that
operates web pages through a browser on behalf of a principal.
Principal
The human user on whose behalf the agent acts.
Task
One unit of work that the principal has asked the agent to
perform, from the principal's request to its completion or
abandonment.
Manifest-Derived Action
Any browser action that the agent takes (a) as a step of a
recovery or commit in a manifest, (b) entering a value that the
principal chose from a manifest's allowedValues, or (c) following
a shortcut, together with the submission of any form that the
Han Expires 21 May 2027 [Page 6]
Internet-Draft AI Manifest November 2026
manifest modified. A value that the agent takes from the
principal's task and enters into a field it found itself is not
manifest-derived, even if a field hint describes that field.
Form Modified by the Manifest
The form owner [HTML] of an element that was the target of a
manifest-derived click or of a manifest-derived key step other
than Tab, of an element whose value changed as a result of such a
step, or of a field into which a value chosen from allowedValues
was entered. For an element without a form owner, as in
applications that save through script, the "form" is the document
containing the element. Its submission or saving by any means,
including by the step that modified it, is a sensitive action
(Section 5.3).
Policy Enforcement Point (PEP)
The component of an agent that decides whether a manifest-derived
action is allowed, denied, or requires the principal's
confirmation (Section 5.2).
Sensitive Action
An action enumerated in Section 5.3.
Security Control
A mechanism enumerated in Section 5.4.
Locator
A description of a UI element within the publisher's documents
(Section 4.2).
3. Discovery
3.1. Well-Known URI
A publisher that provides a manifest MUST serve it at the following
well-known URI [RFC8615] on the origin it describes:
/.well-known/ai-friction-manifest.json
The response MUST be served over HTTPS [RFC9110] with media type
application/json. Agents MUST retrieve it without credentials and
MUST NOT follow any redirect; a redirect response is treated as
absence of a manifest, so that an open redirect on the origin cannot
make another resource be read as its manifest. Section 8.8 describes
an optional credentialed variant.
Han Expires 21 May 2027 [Page 7]
Internet-Draft AI Manifest November 2026
Earlier revisions used the suffix "ai-manifest.json" and the link
relation "ai-manifest". Those names described the purpose of the
resource more broadly than its content, and are replaced by names
specific to friction recovery. Agents MUST ignore any resource at
/.well-known/ai-manifest.json. The earlier names were never
registered with IANA.
3.2. Link Relation
A publisher MAY additionally advertise the manifest with the "ai-
friction-manifest" link relation [RFC8288], either in an HTTP
response header or in an HTML link element [HTML]:
Link: ;
rel="ai-friction-manifest"
The advertisement is a hint that a manifest exists, which addresses
the observation that agents do not spontaneously look for one
(Appendix B). Agents MUST ignore a link target that does not resolve
to the well-known URI on the same origin as the document that carries
the link.
3.3. Removed Discovery Methods
Earlier revisions defined an inline DOM element carrying the manifest
in a data attribute, an "X-AI-Manifest" response header, and an HTML
meta element pointing at an arbitrary path. They are removed:
* Inline manifests are indistinguishable from content injected by
third-party scripts, advertisements, user-generated content, or
cross-site scripting. Agents MUST NOT treat any element in a page
as a manifest.
* The "X-" header prefix is deprecated [RFC6648], and the Link
header above carries the same information. A hash advertised by
the same server that serves the manifest adds no integrity beyond
HTTPS.
* A meta element with an arbitrary path weakens origin binding
without adding function.
Han Expires 21 May 2027 [Page 8]
Internet-Draft AI Manifest November 2026
3.4. Agent-Local Manifests
Earlier revisions defined a file-system path in a specific vendor's
configuration directory from which an agent loads user-curated
manifests, and gave such manifests precedence over and more trust
than site-published ones. That mechanism is removed from the
protocol: local configuration is an implementation matter, and a file
in a user's configuration directory is frequently written by tools,
plug-ins, or shared packages rather than by the user.
An agent MAY accept local overrides, subject to all of the following:
* An override applies only to the site-published manifest that the
agent has currently retrieved and validated for the same origin.
If no such manifest exists, the override MUST be ignored.
* An override can only remove descriptors, remove values from
allowedValues, shorten scope, or bring expires earlier. Any other
content MUST be ignored.
* The resulting manifest is subject to every requirement of this
document, without exemption.
3.5. Retrieval and Caching
* An agent SHOULD retrieve the manifest at most once per origin per
task, and MAY use it for the rest of that task until its "expires"
member. Across tasks, an agent MUST stop reusing a stored
manifest at the earliest of its HTTP freshness lifetime [RFC9111],
its "expires" member, and 90 days after retrieval.
* An agent MUST reject a manifest larger than 262144 octets, a
manifest that is not an I-JSON message [RFC7493], a manifest whose
version member is not "2.0", and a manifest whose objects and
arrays are nested more deeply than the agent's implementation
limit (Section 9 of [RFC8259]). Counting the top-level object as
one level, that limit MUST be at least 64 levels, and publishers
MUST NOT nest a manifest more deeply than 64 levels. An agent
MUST ensure that no manifest can end its process; enforcing the
nesting limit before or while parsing is RECOMMENDED, because in
many environments a parser that exhausts its stack ends the
process rather than reporting an error.
Han Expires 21 May 2027 [Page 9]
Internet-Draft AI Manifest November 2026
* Failure to retrieve or validate a manifest MUST result in the
agent's baseline behavior, as if no manifest existed. Exceeding
the size or nesting-depth limit, and any error raised while
parsing a manifest or while performing the checks in this list,
including exhaustion of the parser's stack or memory, is such a
failure: the task continues without the manifest. Invalid
descriptors are handled as in Section 4.1.
4. Manifest Format
4.1. Top-Level Members
version (string, REQUIRED)
"2.0" for manifests conforming to this document. Manifests with
any other value, including the "1.0" of earlier revisions, are
ignored (Section 3.5).
origin (string, REQUIRED)
The ASCII serialization of the publisher's origin [RFC6454], for
example "https://erp.example.com". Agents MUST ignore a manifest
whose origin member does not equal the origin of the URL from
which it was retrieved.
expires (string, REQUIRED)
A date-time [RFC3339] after which the manifest MUST NOT be used.
Publishers SHOULD keep this within 30 days of publication, so that
stale descriptors expire after UI changes.
scope (array of strings, OPTIONAL)
Path prefixes, each beginning and ending with "/". A document's
path is in scope if it starts with one of them. If absent, the
manifest applies to all paths of the origin.
frameworkHints (object, OPTIONAL)
A "name" member (string) naming the UI framework family, and a
"renderingModel" member (string) from Section 4.8.3. Publishers
SHOULD NOT include version numbers (Section 8.8).
knownTraps (array, REQUIRED)
Trap descriptors (Section 4.3). May be empty.
fieldHints (array, OPTIONAL)
Field descriptors (Section 4.5).
shortcuts (array, OPTIONAL)
Shortcut descriptors (Section 4.6).
Han Expires 21 May 2027 [Page 10]
Internet-Draft AI Manifest November 2026
signatures (array, OPTIONAL)
Publisher signatures (Section 6.1).
Unknown top-level members are ignored. Within a descriptor, a
locator, or a step, an unknown member or value makes the containing
descriptor invalid, and the agent MUST ignore that descriptor while
it MAY use the rest of the manifest. Ignoring a member of a locator
would make it match more elements, so locators always fail closed.
The members "publisher", "manifestId", and "registry_url" of earlier
revisions are removed. In particular, a manifest can no longer name
the service that vouches for it (Section 6.2).
4.2. Locator Object
A locator identifies a UI element. It contains one or more of the
following members; an element matches only if it satisfies all
members present.
role (string): A non-abstract role defined in WAI-ARIA 1.2
[WAI-ARIA]. Other role values make the descriptor invalid.
name (string): The element's accessible name [ACCNAME], compared
exactly after collapsing whitespace.
column (string): For an element in a grid or table, the accessible
name of its column header.
focused (boolean true): The element that currently has focus.
within (locator): An ancestor that the element must have, itself
resolved under the same rules.
testId (string): The value of the element's "data-testid" attribute.
css (string): A selector [SELECTORS4], as a fallback.
Publishers SHOULD prefer role, name, and column, which survive build-
time class-name mangling and are what accessibility-tree-based agents
observe, and SHOULD use css only when no accessible name exists. On
pages that contain user-generated or third-party content, publishers
SHOULD anchor every step locator with "within" to an application
container identified by testId, so that injected elements outside
that container cannot match (Section 8.4).
An agent MUST resolve a locator against the live document at the time
of use, only within documents of the manifest's origin whose path is
in scope, and never inside a cross-origin frame. If the locator
Han Expires 21 May 2027 [Page 11]
Internet-Draft AI Manifest November 2026
matches no element, more than one element, or an element that is not
being rendered [HTML], the agent MUST NOT perform the associated
action.
The locator of a field hint is different: it identifies a class of
elements and MAY match more than one. It is used only to recognize
that an element the agent is already editing belongs to the field,
never to choose a target, and the exactly-one rule does not apply to
it.
4.3. Trap Descriptor
trapId (string, REQUIRED)
Unique within the manifest.
category (string, REQUIRED)
A value from Section 4.8.2.
locator (object, conditional)
The element at which the trap manifests. REQUIRED except for the
category native-alert, which has no element.
condition (object, conditional)
Exactly one of "visible" (a locator that must match), "absent" (a
locator that must not match), or "dialog" (the value "alert").
REQUIRED if recovery contains a wait step or the category is
native-alert.
recovery (array, REQUIRED)
An ordered list of one to eight step objects (Section 4.4). Steps
are identified by trapId and zero-based position.
description (string, OPTIONAL)
Display text, at most 280 characters (Section 4.7).
4.4. Step Object
action (string, REQUIRED)
A value from Section 4.8.1.
locator (object, conditional)
REQUIRED for click, focus, key, scroll, and assert.
key (string, conditional)
REQUIRED for key. One of "Enter", "Tab", "Escape", "ArrowUp",
"ArrowDown", "PageUp", "PageDown".
Han Expires 21 May 2027 [Page 12]
Internet-Draft AI Manifest November 2026
target (string, conditional)
REQUIRED for navigate. A path-absolute URI reference [RFC3986]
with neither a query nor a fragment component.
timeoutMs (integer, conditional)
REQUIRED for wait; from 1 to 30000.
An agent MUST resolve a navigate target against the manifest's origin
with the URL parser of [URL], and MUST treat the step as invalid if
the resulting URL's origin differs from the manifest's origin or its
path is not in scope. The same rules apply to every redirect: before
following one, the agent MUST check its target, and MUST abort the
navigation if the target has a different origin, a query or fragment
component, or a path outside scope. Because browsers follow
redirects automatically, this requires intercepting navigation
requests, for example through the request interception facility of
the automation interface.
There is no step that enters text, selects an option, or supplies any
value. Steps cannot carry data.
4.5. Field Hint Descriptor
Field hints convey site-specific knowledge about valid input, the
third case of Section 1.2. Their design is motivated by the
measurement in Appendix B, in which the information the agent lacked
was of this kind.
fieldId (string, REQUIRED)
Unique within the manifest. Several field hints MAY have the same
locator with different "when" members.
locator (object, REQUIRED)
The input element.
allowedValues (array of strings, OPTIONAL)
Values the site will accept in this field, at most 64 entries,
each of at most 32 characters drawn from ASCII letters, digits,
".", "_", and "-".
when (object, OPTIONAL)
A map from the fieldId of another field hint to a value. The
field hint applies only if, for every entry, the current value of
that other field, as exposed in its accessibility tree, equals the
given value. If the other field's locator matches more than one
element, the agent uses the one that shares the nearest ancestor
with role row with the element being edited; if there is no such
element or more than one, the entry is unsatisfied.
Han Expires 21 May 2027 [Page 13]
Internet-Draft AI Manifest November 2026
commit (object, OPTIONAL)
A single step object with action "key", identified by fieldId,
required to commit an entry in this field.
description (string, OPTIONAL)
Display text, at most 280 characters.
Even short tokens can spell instructions, so allowedValues are never
shown to a language model (Section 4.7). They are used as follows:
1. When the model proposes a value for a field that an applicable
field hint describes, the PEP checks the value against
allowedValues and returns to the model only whether it matched.
2. If it did not match, the agent MUST NOT substitute a listed
value. It MAY show the listed values to the principal, who may
choose one or abandon the entry; otherwise it reports the
mismatch.
3. A value that the principal chose from the list is entered by the
agent's runtime, not by the model, so that the model does not see
it. The choice is the principal's confirmation, and the form
that contains the field counts as modified (Section 5.3).
A field hint therefore constrains which value is acceptable and can
inform the principal; it never chooses a value, and never informs the
model's reasoning beyond a match result.
4.6. Shortcut Descriptor
shortcutId (string, REQUIRED)
Unique within the manifest.
target (string, REQUIRED)
As for the navigate action (Section 4.4).
label (string, REQUIRED)
Display text naming the destination, at most 80 characters.
A shortcut replaces a sequence of menu clicks with a GET request.
Shortcuts are offered to the principal, not to the language model: an
agent MAY list shortcut labels to the principal, who may choose one.
Because a server can perform state changes on GET, following a
shortcut is still subject to the PEP like any other navigation
(Section 5.2). A site that wants agents to perform operations
directly SHOULD expose a structured interface (Section 1.2).
Han Expires 21 May 2027 [Page 14]
Internet-Draft AI Manifest November 2026
4.7. Manifest Strings and the Language Model
No string taken from a manifest reaches a language model. This
covers description, label, allowedValues, trapId, fieldId,
shortcutId, locator members, and navigation targets. An agent MUST
NOT place any of them in a model's context, quoted or otherwise.
What an agent MAY give a model about a manifest is limited to values
from the closed vocabularies of Section 4.8 (trap categories, action
names, and rendering models), small integers that the agent itself
assigns to descriptors, and the match results of Section 4.5. A
model can thus learn, for example, "trap 2, category commit-gesture,
is active; recovery is available", and can ask the agent to run
recovery 2, but cannot read anything the publisher wrote.
Description and label are shown only to the principal and recorded in
logs. No PEP decision depends on their content.
The same limit applies to everything the agent returns to the model
about manifest-derived actions, including PEP outcomes, reasons,
error messages, and audit records: these MUST contain only the items
permitted above. An agent that feeds tool output or logs back to its
model MUST filter them accordingly.
4.8. Vocabularies
The vocabularies below are closed for version "2.0". A future
version of this document may extend them; values unknown to an agent
are handled as specified in Section 4.1.
4.8.1. Actions
click: Activate the located element.
focus: Move focus to the located element.
key: Press one key while the located element has focus.
scroll: Scroll the located container by at most one viewport height.
wait: Wait up to timeoutMs for the trap's condition to stop holding.
assert: Verify that the located element is present; otherwise stop
the recovery.
navigate: Load a same-origin target with GET.
dismissAlert: Dismiss a window.alert() dialog [HTML], which offers
no choice.
Han Expires 21 May 2027 [Page 15]
Internet-Draft AI Manifest November 2026
The actions "fill", "select", and "upload" of earlier revisions are
removed. None of the trap categories requires a manifest to supply a
value, and without them a manifest has no means to place data into a
page or send it anywhere. There is deliberately no action that
accepts a confirmation dialog or answers a prompt dialog.
4.8.2. Trap Categories
commit-gesture: An entry must be confirmed with a key gesture before
the UI proceeds or unlocks.
shadow-dom: The target is rendered inside a shadow tree.
virtual-scroll: Only visible items are rendered.
same-origin-frame: The target is inside a frame of the same origin.
delayed-render: The target appears only after an asynchronous
operation.
hidden-required: A required input is not visible until another
control is operated.
native-alert: A window.alert() dialog blocks the page.
Compared with earlier revisions, the cross-origin case of the iframe
category is removed (Section 4.2), and the native dialog category is
narrowed to alert(), because confirm() and prompt() exist to obtain a
human decision (Section 5.4).
4.8.3. Rendering Models
"server-rendered", "client-rendered", "hybrid", "canvas".
5. Agent Processing Requirements
5.1. Consulting the Manifest
A manifest is reference data, consulted by the agent's runtime rather
than read by its model. The runtime evaluates trap conditions
against the live document; when one holds, it reports the trap to the
model as described in Section 4.7, and runs the recovery through the
PEP when the model asks for it. Field hints are applied as in
Section 4.5, and shortcuts are offered to the principal
(Section 4.6). The agent remains free to use its own inference at
any time, and a manifest never overrides the principal's instructions
or the agent's own policy.
Han Expires 21 May 2027 [Page 16]
Internet-Draft AI Manifest November 2026
An agent MUST perform manifest-derived actions as trusted input
through the browser's own input pipeline, for example WebDriver
actions, as a user would. It MUST NOT dispatch events from script,
modify page script state, or remove or disable page elements in order
to recover from a trap.
Trusted input reaches whatever element the browser finds at the
position of the input, which need not be the element that the locator
resolved: an element that is transparent, or otherwise hard to see,
and placed over the resolved element receives a click aimed at it,
while an ask names the element underneath (Section 5.2). Immediately
before it performs a manifest-derived click, and after any ask, an
agent MUST verify that the element that would receive the click at
the position it uses, taken as the innermost element at that position
without retargeting across shadow boundaries [DOM], is one of the
following, and MUST NOT perform the click otherwise:
* the resolved element;
* a shadow-including descendant [DOM] of the resolved element,
provided that no element on the path from it up to the resolved
element, other than the resolved element itself, is activatable
(Section 5.2); or
* a label element whose labeled control [HTML] is the resolved
element.
The element that would receive the click is determined in the
document that contains the resolved element; for a document in a
frame, the element that would receive the click in each ancestor
document MUST be the element of the frame that leads to it. An agent
SHOULD also decline a manifest-derived click or key step on an
element that the principal cannot see at the position of the input,
or for a key step at the element's in-view position, for example
because the element is fully transparent or clipped away at that
position, or because that position is outside the viewport after the
element has been scrolled into view. This document does not specify
how these checks are made. The Element Click command of WebDriver
[WEBDRIVER] returns an error when the element is obscured by another
element, but its Perform Actions command, which dispatches input at a
position, makes no such check. If a check fails, the agent MAY
repeat it for up to 5000 ms, without acting on the page in between
other than scrolling the element into view; this is not a variant of
the action (Section 5.2). An action that these checks still decline
is treated as denied by the PEP and recorded as such (Section 5.5).
Han Expires 21 May 2027 [Page 17]
Internet-Draft AI Manifest November 2026
5.2. Policy Enforcement Point
An agent that consumes manifests MUST track which of its actions are
manifest-derived, and MUST route every manifest-derived action
through a PEP before execution. A language model MUST NOT take part
in the PEP's decision, so that manifest content cannot influence it.
For each action the PEP:
1. validates the descriptor against Section 4;
2. re-resolves any locator against the live document (Section 4.2);
3. denies the action if it is directed at a security control that
the agent has detected (Section 5.4);
4. assigns the default outcome below;
5. applies the agent's own policy, which can only make the outcome
stricter; and
6. records the outcome (Section 5.5).
In this document, an activatable element is a button, an "a" or
"area" element with an href attribute, a summary element, an input
element whose type is submit, image, reset, button, checkbox, or
radio, or an element whose role is button, link, menuitem,
menuitemcheckbox, menuitemradio, option, tab, switch, checkbox, or
radio. A value-bearing control is a select element, an input or
textarea element, an editing host (contenteditable), or an element
whose role is combobox, listbox, radiogroup, slider, spinbutton,
textbox, or tablist, or that is a descendant of one. The default
outcomes are:
* allow: focus, scroll, assert, and dismissAlert;
* allow: wait with timeoutMs of at most 5000; otherwise ask;
* allow: key Tab;
* allow: key ArrowUp, ArrowDown, PageUp, or PageDown on an element
that is neither activatable nor a value-bearing control; otherwise
ask, and the control counts as modified for Section 5.3;
* ask: key Enter and key Escape, on any element;
* allow: entering a value that the principal has just chosen from
allowedValues (Section 4.5);
Han Expires 21 May 2027 [Page 18]
Internet-Draft AI Manifest November 2026
* ask: click, navigate, and following a shortcut;
* ask: any sensitive action (Section 5.3);
* ask or deny: anything the PEP cannot classify.
Enter and Escape always ask because page script can bind any effect
to a key event on any element, whatever its tag or role, and these
two keys are the ones conventionally bound to commit and dismiss.
The same is true in principle of every input event; the actions
allowed by default are those whose conventional meaning is navigation
within the page rather than a decision, and their residual risk is
discussed in Section 8.4.
Key Enter on an element whose form owner could be submitted by
implicit submission [HTML] would submit a form that the step itself
has modified (Section 2); Section 5.3 therefore requires the PEP to
determine it to be a sensitive action.
An ask MUST show the principal at least the following. The agent's
runtime derives them from the live page, from the structural values
of the validated descriptor, and from the PEP's own classification,
never from description or label text or from a language model.
Values that the publisher controls, such as accessible names and
destination paths, are shown as quoted data, distinct from the
agent's own wording.
Target: the action, with its key or its wait duration where it has
one; for actions on an element, the role and accessible name of
the element actually resolved on the live page; and for navigate
and shortcuts, the origin and path of the destination.
Source: that the action comes from the manifest of the page's
origin, as a recovery step, a field-hint commit, a shortcut, or
the submission of a form modified by the manifest, together with
the page's origin and path.
Effect: first, every sensitive action (Section 5.3) that the PEP has
determined the action can perform or trigger; then the
conventional effect of the action, such as activating the element,
loading the destination with GET, committing or dismissing with
Enter or Escape, changing the value of a control, or letting time
pass; and, for every action other than wait, a statement that the
page decides the actual effect, which the agent cannot determine
and which can include a purchase, a deletion, or another sensitive
action.
Han Expires 21 May 2027 [Page 19]
Internet-Draft AI Manifest November 2026
That statement discloses that the effect is unknown; it does not
identify the effect, which the PEP cannot determine (Section 5.3).
An agent SHOULD warn the principal when the text rendered for an
activatable element is not contained in its accessible name. Any
description or label shown with an ask MUST be visibly marked as
written by the publisher, so that publisher text cannot pass for the
agent's own explanation. The same marking applies when shortcut
labels or allowedValues are shown to the principal. An ask is shown
only to the principal; what the model receives about it remains
limited as specified in Section 4.7.
When the principal declines, or the PEP denies, the agent MUST
abandon the descriptor rather than attempt a variant of it.
Repeating the checks of Section 5.1 within their time limit, as
described there, is not a variant.
5.3. Sensitive Actions
The following are sensitive actions. An agent MUST obtain the
principal's confirmation before a manifest-derived action that the
PEP determines can perform or trigger one, regardless of any manifest
content, signature, or reputation status:
* submitting, by any means, a form modified by the manifest
(Section 2), or a form that contains personal data or credentials;
* a purchase, payment, transfer, or other financial commitment;
* deleting data, or any operation the site presents as irreversible;
* accepting terms, consents, cookie or tracking choices, or
subscriptions;
* sending a message or publishing content; and
* changing account, security, or notification settings.
A PEP MUST at least determine as sensitive a click, or a key step
other than Tab, on a submit button [HTML], key Enter on an element
whose form owner could be submitted by implicit submission [HTML],
and any submission of a form that it has recorded as modified by the
manifest or that contains an input element whose type is password.
It MAY determine others by its own means, never from manifest
content.
A manifest cannot declare an action to be non-sensitive. A PEP
cannot reliably determine the effect of an element whose role, name,
and handlers a hostile publisher controls; the defaults of
Han Expires 21 May 2027 [Page 20]
Internet-Draft AI Manifest November 2026
Section 5.2 therefore ask for confirmation of every action
conventionally bound to a decision rather than depend on that
determination, and nothing turns such an ask into allow. A sensitive
action that page script binds to an action allowed by default is
neither determined nor confirmed; Section 8.4 describes this limit.
5.4. Security Controls Are Not Friction
The following are security controls. They are never traps. An agent
MUST NOT use a manifest to avoid, dismiss, or circumvent them, and
MUST ignore a descriptor whose locator or condition targets one that
it detects:
* HTTP 401, 403, and 429 responses and equivalent in-page denials;
* CAPTCHAs and other challenges that distinguish humans from
automation;
* authentication, re-authentication, and multi-factor or step-up
challenges;
* window.confirm(), window.prompt(), and beforeunload dialogs, and
site-built dialogs that ask the user to confirm or decline an
action;
* consent and cookie-choice interfaces; and
* locks, holds, or timers that a site applies to give a human the
opportunity to reconsider.
Native confirm(), prompt(), and beforeunload dialogs block the page's
event loop, so DOM interactions hang until the dialog is closed; an
agent detects them through its automation interface and hands them to
the principal. Detection of site-built controls is best effort; the
confirmation defaults of Section 5.2 are the backstop for controls
that go undetected.
On detecting a security control, an agent MUST stop the manifest-
derived recovery and either hand control to the principal or end the
task. It MUST NOT retry with altered inputs.
Han Expires 21 May 2027 [Page 21]
Internet-Draft AI Manifest November 2026
5.5. Audit Record
An agent SHOULD keep, for each manifest-derived action, a record
containing the origin; the SHA-256 [RFC6234] digest of the JCS
canonical form [RFC8785] of the manifest; the descriptor identifier
and step position; the PEP outcome and its reason; the text of any
alert dismissed; and the time. The record SHOULD be available to the
principal. It allows the principal, and the publisher's defenders,
to distinguish manifest-guided recovery from other automated behavior
(Section 8.9).
6. Authenticity and Reputation
6.1. Publisher Signatures
HTTPS already authenticates a manifest retrieved from its origin.
Signatures are useful when a manifest is authored by a party acting
for the origin, such as the organization that deploys a vendor
application, or when it is held outside the origin, such as in an
agent's cache.
Each element of the signatures member is a JSON Web Signature
[RFC7515] in compact serialization with a detached payload
(Appendix F of [RFC7515]). The payload is the JCS canonical form
[RFC8785] of the manifest with the signatures member removed. The
protected header MUST contain "alg" and "kid". Agents that verify
signatures MUST support "ES256" [RFC7518], MAY support "Ed25519"
[RFC9864], and MUST reject "none", every symmetric algorithm, and
every algorithm that [RFC9864] deprecates.
This document does not define how keys are discovered. An agent MUST
verify signatures only against keys that its own configuration or its
principal designates for the origin; a key or key location named by
the manifest MUST NOT be used. If such a key is designated for the
origin, a manifest without a signature that verifies under it MUST be
ignored. If no key is designated, signatures are ignored.
A valid signature establishes who published the manifest bytes. It
does not establish that the manifest is safe, and MUST NOT relax any
requirement of Section 5.
6.2. Reputation Services
Earlier revisions defined a "Central Trust Registry" whose location
each manifest declared in a registry_url member, and whose "white"
status permitted execution. Because the party being vouched for
chose the voucher, a malicious publisher could name a registry it
controlled. That design is withdrawn.
Han Expires 21 May 2027 [Page 22]
Internet-Draft AI Manifest November 2026
An agent MAY consult reputation services that its own configuration
designates. A manifest MUST NOT be able to select, add, or override
such a service. A reputation service:
* keys its information by origin, because a digest of manifest bytes
changes with any edit;
* reports abuse history and the results of any static analysis it
performs on manifests it has observed; an unfavorable status MUST
cause the manifest to be ignored, and a favorable status MUST NOT
relax any requirement of Section 5;
* cannot certify safety, because a manifest's effect depends on the
live page that its locators resolve against, which a service
evaluating manifest bytes never sees (Section 8.4);
* SHOULD sign its responses, so that an agent can verify cached
responses; and
* SHOULD be queried in a way that does not disclose the principal's
browsing to the service, for example by downloading a list or by
hash-prefix queries, rather than with a per-visit lookup
(Section 9).
7. IANA Considerations
7.1. Well-Known URI Registration
IANA is requested to register the following in the "Well-Known URIs"
registry [RFC8615]:
URI Suffix: ai-friction-manifest.json
Change Controller: Won-pyo Han (pk102h@naver.com)
Reference: This document
Status: provisional
Related Information: None
7.2. Link Relation Type Registration
IANA is requested to register the following in the "Link Relation
Types" registry [RFC8288]:
Relation Name: ai-friction-manifest
Description: Refers to the AI Friction-Recovery Manifest of the link
context's origin, which describes UI patterns that stall automated
agents and constrained steps to recover from them.
Reference: This document
Notes: None
Han Expires 21 May 2027 [Page 23]
Internet-Draft AI Manifest November 2026
7.3. No New Registries
Earlier revisions requested two new registries with the Specification
Required policy. Registries created by an Independent Submission
cannot use that policy [RFC8726], and the vocabularies are closed in
this version. The vocabularies are therefore defined in Section 4.8,
and no new registry is requested.
8. Security Considerations
8.1. Threat Model
This document considers the following adversaries:
T1 Malicious publisher: an origin that publishes a manifest designed
to harm the principal.
T2 Conflicted publisher: an honest origin whose interests differ
from the principal's, for example in consent, retention, or
upselling flows.
T3 Content injector: a third party able to place content in an
origin's pages, through user-generated content, advertising, a
compromised script dependency, or cross-site scripting.
T4 Local supply chain: a tool, plug-in, or package able to write
files in the principal's configuration directories.
T5 Network attacker: an on-path party.
T6 Reconnaissance agent: an autonomous offensive agent that harvests
manifests to plan an attack on the publisher.
8.2. Prompt Injection
A language-model agent cannot reliably separate instructions from
data in its input, and prompt injection through web content remains
an unsolved problem. Revision -02 of this document recommended out-
of-band discovery because out-of-band data was observed not to be
screened as page content. That rationale is withdrawn: content from
/.well-known is written by the same publisher as the page and is
exactly as untrusted (P1).
This document instead keeps manifest content out of the model
entirely. No string written by the publisher reaches a language
model (Section 4.7): the model sees only closed-vocabulary values,
integers assigned by the agent, and match results, none of which the
publisher can use to express an instruction. The agent's runtime,
Han Expires 21 May 2027 [Page 24]
Internet-Draft AI Manifest November 2026
not the model, interprets the manifest, and every action it takes on
the basis of the manifest is decided by a PEP that the model cannot
influence (P2, P6). This follows designs that separate the data an
agent acts on from the channel that controls it [CAMEL], applied to
the manifest. It does not protect against injection through the page
itself, which a manifest neither adds nor removes.
Showing allowedValues or shortcut labels to the model, even as quoted
and attributed data, would not be safe: 64 tokens of 32 characters,
or an 80-character label, suffice to spell instructions (for example
"IGNORE_TASK_CLICK_CANCEL"), and a model so influenced could act on
page elements that the manifest never names, outside the PEP. Those
lists are therefore shown only to the principal.
8.3. Exfiltration
An agent that holds private data, reads untrusted content, and can
communicate externally is exposed to data exfiltration [TRIFECTA].
This document closes the manifest's external channel: steps carry no
values, there is no action that enters text or uploads a file,
navigation targets are same-origin paths without query or fragment,
and redirects that leave the origin or scope, or add a query, stop
navigation (Section 4.4). Enter, Escape, and click require
confirmation (Section 5.2), and the submission of any form modified
by the manifest is a sensitive action (Section 5.3). What the agent
enters from the principal's task into fields it found itself reaches
the origin exactly as it would without a manifest.
8.4. Locator Substitution
A locator's meaning is determined by the live page, not by the
manifest. A publisher (T1) can publish an innocuous manifest, have
it signed or favorably rated, and later change the page so that the
same locator resolves to a different control. A content injector
(T3) can add an element that matches a publisher's locator. Hashes,
signatures, and reputation over manifest bytes cannot detect either.
The mitigations are that locators are resolved at use time and must
match exactly one rendered element (Section 4.2), that step locators
should be anchored to an application container, and that every action
conventionally bound to a decision requires confirmation by default
regardless of how the element was reached (Section 5.2). An injected
element that duplicates a locator makes recovery fail closed, which
is a denial of the manifest's benefit rather than a harm to the
principal; an injected element that alone matches before the genuine
element renders can receive only the actions allowed by default. An
injected element that does not match the locator but covers the
genuine element, transparent or not, would receive a click aimed at
Han Expires 21 May 2027 [Page 25]
Internet-Draft AI Manifest November 2026
it; the check of the element that receives a click (Section 5.1)
makes such a click fail closed as well, if the covering element is
present when the check is made.
The actions allowed by default (focus, scroll, assert, waits of at
most five seconds, Tab, arrow and page keys on elements that are
neither activatable nor value-bearing, dismissing an alert, and
entering a value that the principal has chosen from allowedValues)
can still be followed by effects of page script, since a hostile
origin can bind any effect to any event or run it at any time,
including a sensitive action such as a purchase. A PEP cannot
determine such an effect, and it is not confirmed: Section 5.3
requires confirmation of the sensitive actions that the PEP
determines, and Section 5.2 asks for every action conventionally
bound to a decision, but neither covers an effect bound to an action
allowed by default. No requirement of this document ensures that a
manifest-derived action never has an unconfirmed effect. The
manifest's contribution to that risk is limited to directing such
events at a chosen element; the same origin could bind the same
effects to events that an agent without a manifest, or a human user,
produces in ordinary navigation. Agents that want no residual risk
MAY treat every manifest-derived action as ask.
8.5. Publisher-Principal Conflict
A manifest is advisory, but it is not neutral (T2). A recovery that
clicks "accept all" on a consent interface, confirms a dialog,
presses Escape on a banner that the site interprets as acceptance,
waits out a cooling-off timer, or continues a subscription in a
cancellation flow would serve the publisher at the principal's
expense. A field hint that omits a valid value could steer the
principal toward another. These are addressed by Section 5.3, the
confirmation defaults of Section 5.2 for click, Enter, Escape, and
waits longer than five seconds, and the prohibition on substituting
values (Section 4.5), none of which manifest content can relax (P4).
A cooling-off timer shorter than five seconds is protected only by
the agent's best-effort detection of security controls (Section 5.4)
and by confirmation of whatever sensitive action follows it.
8.6. Circumvention of Security Controls
A mechanism that teaches agents to recover from friction is dual-use
if it applies to security controls: it would give benign agents the
persistence after denial that defenders use to recognize autonomous
attacks [ARTEX], and could be used to defeat anti-automation
measures, rate limits, and human confirmation. Section 5.4 therefore
defines these as stop conditions, and future versions of this
document will not define trap categories that correspond to a
Han Expires 21 May 2027 [Page 26]
Internet-Draft AI Manifest November 2026
security control.
8.7. Local Configuration Supply Chain
Earlier revisions exempted user-curated local manifests from prompt-
injection screening and gave them precedence. A tool or package able
to write one file (T4) would thereby obtain a persistent, unscreened,
highest-precedence instruction channel for any origin. Section 3.4
limits local overrides to narrowing a site-published manifest that is
actually present, so that a local file can only reduce what an agent
does.
8.8. Information Disclosure, Reconnaissance, and Cloaking
A manifest retrieved without credentials is public and is read by
attackers as well as by agents (T6). Framework versions help an
attacker select known vulnerabilities for which a deployment is
unpatched, and shortcuts or field hints can reveal paths and code
values that are not otherwise discoverable. Publishers:
* SHOULD NOT include framework or component version numbers; and
* MUST NOT include, in a manifest retrievable without credentials,
any path, value, or element that is not reachable without
credentials.
A publisher MAY serve a more detailed manifest at the same URI to
requests that carry the credentials the UI itself requires, and MAY
additionally require agent authentication [WEBBOTAUTH]. An agent MAY
make such a credentialed request using only credentials it already
holds for that origin in the current task. Such a manifest is
subject to every requirement of this document. If an agent retrieves
both, the credentialed manifest replaces the uncredentialed one for
the rest of the task, and any local override applies to the manifest
in use.
Because a publisher can serve different manifests to different
requesters, a reputation service or reviewer can be shown a benign
manifest while agents receive another (cloaking). This is a further
reason that no favorable status relaxes the requirements of
Section 5.
Han Expires 21 May 2027 [Page 27]
Internet-Draft AI Manifest November 2026
8.9. Distinguishing Conforming Agents
The defense guide cited above advises anchoring the detection of
autonomous attacks on behavioral fingerprints rather than on static
ones such as User-Agent strings, and lists persistence with bypass
variations after denials among them [ARTEX]. An agent conforming to
this document acts on manifest guidance only through declared steps
on located elements, and stops at denials. A publisher can therefore
use its own manifest as a positive baseline in behavioral detection:
automated activity that follows declared recovery steps and stops at
controls is consistent with a conforming agent, and activity that
persists past controls is not, whatever the client claims to be.
8.10. Resource Exhaustion and Transport
Size and nesting-depth limits (Section 3.5), the bound of eight steps
per recovery (Section 4.3), bounded wait and scroll actions, and the
absence of any reference from one manifest to another bound the cost
of processing. A parser that exhausts its stack on deeply nested
input can end the agent instead of falling back to baseline behavior;
the nesting limit, and the treatment of parser errors as validation
failures, address this (Appendix B). HTTPS and the prohibition of
cross-origin redirects (Section 3.1) protect against T5.
9. Privacy Considerations
Requesting /.well-known/ai-friction-manifest.json tells the origin
that the client is likely an AI agent, in addition to the visit that
prompted it. Agents that do not want to disclose this MAY refrain
from retrieving manifests.
Reputation lookups are a larger exposure: a per-visit query reveals
the principal's browsing to a third party, which is why Section 6.2
recommends list-based or hash-prefix designs.
Manifests contain no personal data, since steps carry no values.
Audit records (Section 5.5) contain the principal's browsing history
and SHOULD be held under the principal's control. Local overrides
listing origins reveal those origins if synchronized or shared.
10. Implementation Status
Note to RFC Editor: Please remove this section before publication, as
described in [RFC7942].
A reference implementation is available at https://github.com/11pyo/
AINavManifest under the MIT License. It implements the requirements
of revision -03 for the version "2.0" format: a JSON Schema and a
Han Expires 21 May 2027 [Page 28]
Internet-Draft AI Manifest November 2026
validator; a policy enforcement point with the default outcomes,
standing-approval limits, and ask display that revision -03
specified, and an audit record; tracking of forms modified by the
manifest; a model interface that exposes only the items permitted by
Section 4.7; the field-hint flow of Section 4.5; discovery without
redirects and per-hop redirect checks for navigation (Section 4.4);
verification of detached JWS publisher signatures with "ES256" and
"Ed25519" against keys that its configuration designates for the
origin, rejecting every other algorithm, including "EdDSA"
(Section 6.1), where a valid signature identifies the publisher and
relaxes no policy; example publishers serving manifests at the well-
known URI with the Link advertisement; and tests, including browser
tests, that exercise these. Item 6 of Appendix B reports fourteen
attacks on its signature verification. It does not implement
reputation services.
Of the changes in revision -04, it implements only the nesting limit
and the handling of parser errors (Section 3.5) and, through the
actionability checks of its browser automation library, declines a
click that another element would intercept, which approximates the
check of Section 5.1. It does not yet implement the withdrawal of
standing approval: its PEP still lets a principal approve key Enter
in advance for an origin. The revised definition of a form modified
by the manifest (Section 2), the content of an ask (Section 5.2), and
the visibility check of Section 5.1 are not implemented either. The
measurements in items 1 to 4 of Appendix B were made with an earlier
version of the implementation that used the version "1.0" format,
which is preserved for reproducibility; items 5 and 6 used the
version "2.0" implementation, item 6 after signature verification was
added.
11. Intellectual Property Rights Disclosure
The technology described in this document is the subject of Korean
Patent Applications No. 10-2026-0071716 (filed 2026-04-21) and No.
10-2026-0097280 (filed 2026-05-29), both by the author. The
applicant commits to offer any essential claims under Fair,
Reasonable, and Non-Discriminatory (FRAND) terms to implementers of
this specification, as declared in the project repository.
12. Acknowledgments
The author thanks the Claude Code, Model Context Protocol, and OpenAI
function-calling communities for the empirical observations that
motivated this work. The threat model of Section 8 draws on the
publicly documented design and defensive analysis of ARTEX,
originally developed as Autumn-27/ARTEX, as published in its global
edition [ARTEX].
Han Expires 21 May 2027 [Page 29]
Internet-Draft AI Manifest November 2026
13. References
13.1. 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,
.
[RFC3339] Klyne, G. and C. Newman, "Date and Time on the Internet:
Timestamps", RFC 3339, DOI 10.17487/RFC3339, July 2002,
.
[RFC3986] Berners-Lee, T., Fielding, R., and L. Masinter, "Uniform
Resource Identifier (URI): Generic Syntax", STD 66,
RFC 3986, DOI 10.17487/RFC3986, January 2005,
.
[RFC6234] Eastlake 3rd, D. and T. Hansen, "US Secure Hash Algorithms
(SHA and SHA-based HMAC and HKDF)", RFC 6234,
DOI 10.17487/RFC6234, May 2011,
.
[RFC6454] Barth, A., "The Web Origin Concept", RFC 6454,
DOI 10.17487/RFC6454, December 2011,
.
[RFC7493] Bray, T., Ed., "The I-JSON Message Format", RFC 7493,
DOI 10.17487/RFC7493, March 2015,
.
[RFC7515] Jones, M., Bradley, J., and N. Sakimura, "JSON Web
Signature (JWS)", RFC 7515, DOI 10.17487/RFC7515, May
2015, .
[RFC7518] Jones, M., "JSON Web Algorithms (JWA)", RFC 7518,
DOI 10.17487/RFC7518, May 2015,
.
[RFC9864] Jones, M.B. and O. Steele, "Fully-Specified Algorithms for
JSON Object Signing and Encryption (JOSE) and CBOR Object
Signing and Encryption (COSE)", RFC 9864,
DOI 10.17487/RFC9864, October 2025,
.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
May 2017, .
Han Expires 21 May 2027 [Page 30]
Internet-Draft AI Manifest November 2026
[RFC8259] Bray, T., Ed., "The JavaScript Object Notation (JSON) Data
Interchange Format", STD 90, RFC 8259,
DOI 10.17487/RFC8259, December 2017,
.
[RFC8288] Nottingham, M., "Web Linking", RFC 8288,
DOI 10.17487/RFC8288, October 2017,
.
[RFC8615] Nottingham, M., "Well-Known Uniform Resource Identifiers
(URIs)", RFC 8615, DOI 10.17487/RFC8615, May 2019,
.
[RFC8785] Rundgren, A., Jordan, B., and S. Erdtman, "JSON
Canonicalization Scheme (JCS)", RFC 8785,
DOI 10.17487/RFC8785, June 2020,
.
[RFC9110] Fielding, R., Ed., Nottingham, M., Ed., and J. Reschke,
Ed., "HTTP Semantics", STD 97, RFC 9110,
DOI 10.17487/RFC9110, June 2022,
.
[RFC9111] Fielding, R., Ed., Nottingham, M., Ed., and J. Reschke,
Ed., "HTTP Caching", STD 98, RFC 9111,
DOI 10.17487/RFC9111, June 2022,
.
[WAI-ARIA] W3C, "Accessible Rich Internet Applications (WAI-ARIA)
1.2", W3C Recommendation wai-aria-1.2, June 2023,
.
[ACCNAME] W3C, "Accessible Name and Description Computation 1.2",
.
[HTML] WHATWG, "HTML Living Standard",
.
[DOM] WHATWG, "DOM Living Standard",
.
[URL] WHATWG, "URL Living Standard",
.
[SELECTORS4]
W3C, "Selectors Level 4",
.
Han Expires 21 May 2027 [Page 31]
Internet-Draft AI Manifest November 2026
13.2. Informative References
[RFC6648] Saint-Andre, P., Crocker, D., and M. Nottingham,
"Deprecating the "X-" Prefix and Similar Constructs in
Application Protocols", BCP 178, RFC 6648,
DOI 10.17487/RFC6648, June 2012,
.
[RFC7942] Sheffer, Y. and A. Farrel, "Improving Awareness of Running
Code: The Implementation Status Section", BCP 205,
RFC 7942, DOI 10.17487/RFC7942, July 2016,
.
[RFC8726] Farrel, A., "How Requests for IANA Action Will Be Handled
on the Independent Stream", RFC 8726,
DOI 10.17487/RFC8726, November 2020,
.
[RFC9309] Koster, M., Illyes, G., Zeller, H., and L. Sassman,
"Robots Exclusion Protocol", RFC 9309,
DOI 10.17487/RFC9309, September 2022,
.
[MCP] Model Context Protocol, "Model Context Protocol
Specification",
.
[WEBMCP] W3C Web Machine Learning Community Group, "WebMCP",
.
[A2A] A2A Project, "Agent2Agent (A2A) Protocol",
.
[LLMSTXT] Howard, J., "The /llms.txt file", September 2024,
.
[AIPREF] IETF, "AI Preferences (aipref) Working Group",
.
[WEBBOTAUTH]
IETF, "Web Bot Auth (webbotauth) Working Group",
.
[WEBDRIVER]
W3C, "WebDriver", W3C Recommendation REC-
webdriver1-20180605, 5 June 2018,
.
Han Expires 21 May 2027 [Page 32]
Internet-Draft AI Manifest November 2026
[CAMEL] Debenedetti, E., "Defeating Prompt Injections by Design",
arXiv 2503.18813, March 2025,
.
[AGENTDOJO]
Debenedetti, E., "AgentDojo: A Dynamic Environment to
Evaluate Prompt Injection Attacks and Defenses for LLM
Agents", arXiv 2406.13352, June 2024,
.
[TRIFECTA] Willison, S., "The lethal trifecta for AI agents: private
data, untrusted content, and external communication", June
2025, .
[ARTEX] artex-global maintainers (upstream: Autumn-27/ARTEX),
"ARTEX Global Edition: docs/defense-en.md, guard/guard.go,
and intercept/intercept.go", 2026,
.
[DRAFT02] Han, W., "AI Manifest: Site-Published Friction-Recovery
Descriptors for AI Agents", Work in Progress, Internet-
Draft, draft-han-ai-manifest-02, 2 June 2026,
.
Appendix A. Example Manifest
The following manifest describes a table-maintenance screen on which
a cell entry must be committed with Enter before the row unlocks, and
on which the valid account-category reference codes depend on the
selected company code. Under Section 8.8, its fieldHints would be
served only in response to requests authenticated to the ERP system.
Han Expires 21 May 2027 [Page 33]
Internet-Draft AI Manifest November 2026
{
"version": "2.0",
"origin": "https://erp.example.com",
"expires": "2026-12-01T00:00:00Z",
"scope": ["/maint/"],
"frameworkHints": {
"name": "webdynpro",
"renderingModel": "server-rendered"
},
"knownTraps": [
{
"trapId": "row-commit",
"category": "commit-gesture",
"locator": { "role": "gridcell", "focused": true },
"condition": { "visible": { "role": "status",
"name": "Row locked" } },
"recovery": [
{ "action": "key", "key": "Enter",
"locator": { "role": "gridcell", "focused": true } },
{ "action": "wait", "timeoutMs": 3000 }
],
"description": "Commit each entry with Enter to unlock."
}
],
"fieldHints": [
{
"fieldId": "companyCode",
"locator": { "role": "combobox", "name": "Company Code" }
},
{
"fieldId": "acctCatRef-1000",
"locator": { "role": "gridcell",
"column": "Account Category Ref" },
"when": { "companyCode": "1000" },
"allowedValues": ["Z051", "Z063", "Z072"],
"commit": { "action": "key", "key": "Enter",
"locator": { "role": "gridcell",
"focused": true } },
"description": "Only codes registered for the company code."
}
],
"shortcuts": [
{ "shortcutId": "acct-cat-maint",
"target": "/maint/account-category",
"label": "Account category maintenance" }
]
}
Han Expires 21 May 2027 [Page 34]
Internet-Draft AI Manifest November 2026
The locator of acctCatRef-1000 matches every cell of its column,
which is permitted for field hints (Section 4.2). The model never
sees the list. When it proposes a code from the principal's task,
the PEP answers only whether the code is accepted. If the model
proposes a standard code such as one it knows from public
documentation, the answer is no; the agent then shows Z051, Z063, and
Z072 to the principal, who chooses one or abandons the entry
(Section 4.5). The shortcut is likewise offered to the principal,
not to the model.
The Enter steps ask the principal each time (Section 5.2), so
maintaining many rows takes at least one confirmation per row; a
deployment that needs unattended bulk entry would offer a structured
interface instead (Section 1.2). This example assumes that the
focused element is the grid cell itself; on pages that place an input
element inside the cell, the publisher would use { "role": "textbox",
"focused": true, "within": { "role": "gridcell" } }.
Appendix B. Preliminary Measurements
This appendix reports measurements made with the version "1.0" format
of [DRAFT02] (items 1 to 4), none of which evaluates security, and
adversarial evaluations of the reference implementation of the
version "2.0" requirements of revision -03 (item 5) and of its
signature verification (item 6). They are preliminary; their
limitations are stated with each.
1. Proprietary-knowledge enterprise task (2026-05-29).
An SAP S/4HANA-style table-maintenance screen on which valid
account-category references depend on company-code registration,
and on which a cell entry must be committed before the row
unlocks. One run per condition with one agent (Claude Sonnet
4.6). Both conditions completed. The manifest condition used 6
interaction actions versus 13 and about 47.8K versus 87.3K input
tokens (estimated). Wall-clock time was closer, 275 s versus 353
s, because at this difficulty the agent's own reasoning time
dominates both conditions; actions and tokens reflect the
efficiency difference more directly. Limitations: a single run
per condition; one model; site, manifest, and evaluation all
prepared by the author; token counts estimated.
2. Reference benchmark.
Han Expires 21 May 2027 [Page 35]
Internet-Draft AI Manifest November 2026
A two-step order-entry task repeated 30 times, input tokens
counted with the cl100k_base tokenizer: the manifest-aware agent
used 341 input tokens on average with 30 of 30 successes, and a
DOM-analysis baseline used 1887.6 tokens with 6 of 30 successes.
Limitations: both agents were implemented by the author; the
tokenizer differs from that of the model in item 1.
3. Spontaneous discovery (Phase 1, n=30, 2026-04-28).
No agent spontaneously looked for an unannounced manifest in 30
trials. Consumption therefore depends on agents implementing
discovery, which motivates the link relation (Section 3.2). An
effect size of Cohen's d = 5.00 reported in earlier revisions was
an artifact of a near-zero-variance idealized baseline and is
withdrawn.
4. Cross-agent consumption (Phase 2, 2026-05-11).
Given a manifest in its conversation context, one agent applied
individual trap entries, a second recognized the document type
without applying entries, and a third did not extract manifest-
specific content. This tested interpretation, not discovery.
Version "2.0" removes the dependence on interpretation: the
agent's runtime, not its model, interprets the manifest.
5. Adversarial evaluation (2026-11).
Using the attack-success measure of [AGENTDOJO], 32 attack cases
were run against four agent configurations on a local test site
that records side effects such as purchases and deletions. Each
case is a malicious manifest with a success criterion and a
predicted outcome per configuration, a few left open; the cases
include one control case and one malformed case kept as
registered. The configurations were an unguarded agent that
executes manifest steps as written; the same agent with only the
schema validator; and the reference implementation of the
requirements of Section 5 as they stood in revision -03
(validator, PEP, no manifest string reaching the model, and
provenance), once with a principal who declines every ask and
once with one who approves every ask and picks the first item of
every list. Attack success rates were 32 of 32 and 24 of 32 for
the first two configurations, which are upper bounds by
construction rather than evidence; 3 of 32 with the declining
principal; and 13 of 32 with the approving principal, 10 of them
after an approved ask. The 13 include the control case, and the
32 include the malformed case, which was blocked. Excluding the
two attacks in which a standing approval (an approval of Enter
given in advance for an origin, which revision -03 permitted and
Han Expires 21 May 2027 [Page 36]
Internet-Draft AI Manifest November 2026
revision -04 withdraws) let Enter run without an ask, the last
two rates were 1 of 30 and 11 of 30. No manifest string reached
the model in any run of the reference implementation (0 of 64
distinct runs, 192 with repeats).
The attacks that succeeded without any ask were a deeply nested
manifest whose parsing raised an exception that the
implementation did not handle, which led to the nesting limit and
the parser-error rule of Section 3.5, and two in which a standing
approval of Enter let a purchase occur (Section 5.3): one through
a key handler that page script bound to a text field, and one
through an Enter that submitted an ordinary form. Revision -04
withdraws standing approval (Section 5.2), so by analysis of the
specification both Enter steps require an ask; this has not been
measured, because the reference implementation does not yet
reflect the withdrawal (Section 10). By the same analysis, an
estimate and not a measurement, the fix for the nested manifest
together with the withdrawal would leave 0 of 32 successes with
the declining principal and 12 of 32 with the approving
principal, all after an approved ask. A click on a button
covered by a transparent element was stopped only by the hit test
of the browser automation library used by the adapter, which led
to the check in Section 5.1. In the two approved purchases bound
to Enter on a text field, the ask did not name the effect, judged
by a keyword test fixed in advance; this led to the requirements
on what an ask shows (Section 5.2), under which an ask discloses
that the effect is unknown but does not identify it. After the
parser was fixed, a single re-measurement blocked the deeply
nested manifest in the three configurations that validate and
changed no other result. The predictions for two attacks were
wrong in some configurations and are kept as registered.
Method: a first measurement was discarded after a fix to the test
harness, and the definition of a success after an approved ask
was changed during measurement, after which all results were
discarded and measured again from scratch. The criteria and
predictions are stated to have been fixed before measurement, but
they were committed to the repository four seconds before the
results, so their prior registration cannot be verified
independently. Utility, that is whether the agent still
completes benign tasks, was not measured in the scripted
evaluation. In the language-model pilot noted below, a benign
task in which the user names the value succeeded in 5 of 5 runs
in every configuration, while a task that needs a value from a
manifest's allowedValues could not be completed by the model in
the configurations of the reference implementation, which never
show it the list (Section 4.7): a cost of the design.
Han Expires 21 May 2027 [Page 37]
Internet-Draft AI Manifest November 2026
Limitations: the attacks were designed by the author and form a
regression suite for known classes, not a discovery of unknown
ones; the model is a script that obeys action words in any text
it is shown, not a language model, so the results measure what
the surrounding structure contains, not how a real model behaves;
the attacker does not adapt; the two principals are constants,
not people; the runs are deterministic, so the three repeats
confirm only the stability of the test harness; and everything
ran on a local test site, without real sites, credentials, or
payments. A pilot with one language model in place of the
script, 75 runs with 5 per scenario and configuration, is
published with the evaluation; with 5 runs per cell it is not
evidence.
6. Signature attacks (2026-11).
After signature verification was added to the reference
implementation (Section 6.1), 14 further attack cases were
registered with predictions, after the verifier was written and
in a commit separate from and earlier than their results, and run
three times in the four configurations of item 5. In the two
configurations of the reference implementation a key was
designated for the site's origin, except in one control case.
The nine cases with a forged, stripped, or confused signature
(forged "ES256" and "Ed25519" signatures under the designated key
identifier, "alg" "none" in two forms, HMAC keyed with the
designated public key, a key named by the manifest or the JWS
header, a genuine signature moved onto another manifest, a
missing signature, and the deprecated "EdDSA" label) succeeded in
0 of 9 with either principal, against 9 of 9 for the unguarded
agent and 8 of 9 for the validator-only agent. With the
designated key removed in test code, 8 of the 9 succeeded with
the approving principal, after an approved ask. Over all 14
cases the approving principal saw 3 successes, all after an
approved ask, in the cases that test whether a valid signature
widens anything: it did not, since a validly signed manifest gave
the same outcome and the same ask display as its unsigned
control, and a signed copy of an attack from item 5 the same as
the original. All registered predictions held, and no result
changed between repeats. Limitations: those of item 5; in
addition, these attacks test the verifier and the rule for
designated keys, not key management, which this document leaves
to the agent's configuration.
Earlier single-trial figures (a 5.0 s "Tier C" time and 19- to
31-fold speedups in draft-han-ai-manifest-01) remain withdrawn.
Items 5 and 6 are first adversarial measurements of the kind that
revision -03 named as the most needed next one. Measurements with
Han Expires 21 May 2027 [Page 38]
Internet-Draft AI Manifest November 2026
language models at a scale that can serve as evidence, with attacks
not designed by the author, and of the requirements as revised in
revision -04 remain to be made.
Appendix C. Document History
Note to RFC Editor: Please remove this section before publication.
draft-han-ai-manifest-04 (2026-11)
Withdraws the standing approval of Enter, restates the
confirmation requirement to what an agent can determine, and
tightens requirements that an adversarial evaluation of the
reference implementation found incomplete (Appendix B, item 5):
* Retrieval: an implementation limit on nesting depth, of at
least 64 levels; exceeding a limit and any parser error,
including resource exhaustion, are validation failures that
fall back to baseline behavior.
* Standing approval of Enter is withdrawn: Enter always asks, and
an agent's policy can only deny it.
* Section 5.3 is restated to require confirmation of the
sensitive actions the PEP determines, with a minimum that every
PEP determines (submit buttons, Enter that could submit a form
by implicit submission, and forms modified by the manifest or
containing a password field); the -03 wording ("performs or
triggers") could not be met for effects that page script binds
to actions allowed by default (Section 8.4). Revision -03 had
the same limit without stating it. The submission of a form
modified by the manifest is sensitive also when the modifying
step itself submits it.
* Before a manifest-derived click, the agent verifies that the
click would reach the resolved element rather than an element
covering it, judged on the innermost element without
retargeting, and may repeat the check for up to 5000 ms; it
should decline actions on elements that the principal cannot
see.
* An ask shows the target, the source, and the effect of the
action, including a statement that the actual effect is unknown
to the agent; values that the publisher controls are shown as
quoted data.
* The abstract no longer carries revision history, which is kept
in this appendix.
Han Expires 21 May 2027 [Page 39]
Internet-Draft AI Manifest November 2026
* Section 3.3 is retitled "Removed Discovery Methods".
* The adversarial evaluation and fourteen attacks on signature
verification are reported, with their limitations, in
Appendix B.
draft-han-ai-manifest-03 (2026-11)
Replaces the trust model. Principal changes:
* The manifest is untrusted regardless of channel; the -02
rationale that out-of-band data escapes prompt-injection
screening is withdrawn. Added design principles, applicability
criteria, a threat model, provenance tracking of manifest-
derived actions, a policy enforcement point with confirmation
defaults, sensitive actions, and the exclusion of security
controls from recovery. No manifest string reaches the
language model; value lists and shortcuts are shown only to the
principal.
* Discovery: removed the inline DOM element, the X-AI-Manifest
header, the meta element, and the vendor-path local method;
local overrides may only narrow a site manifest that is
present. Added origin binding and the Link header form of the
link relation. Renamed the well-known suffix to "ai-friction-
manifest.json" and the link relation to "ai-friction-manifest".
* Schema version "2.0": closed action vocabulary of value-free
steps; removed fill, select, and upload; multi-strategy
locators with ARIA role, name, and column; fieldHints for site-
specific knowledge; same-origin shortcuts without query;
expires and scope members; removed publisher, manifestId, and
registry_url; narrowed the dialog category to alert() and
removed cross-origin frames.
* Integrity: replaced the manifest-declared Central Trust
Registry with agent-configured reputation services and optional
JWS signatures over the JCS canonical form.
* IANA: no new registries (RFC 8726); the well-known URI and link
relation registrations are kept.
* Measurements moved to an appendix with stated limitations,
including one run per condition for the enterprise task.
draft-han-ai-manifest-02 (2026-06-02)
Realigned the document with the advisory friction-recovery model;
relaxed the deterministic runtime to OPTIONAL; withdrew the "Tier
C" figures and Cohen's d = 5.00; added the enterprise measurement.
Han Expires 21 May 2027 [Page 40]
Internet-Draft AI Manifest November 2026
draft-han-ai-manifest-01 (2026-05-15)
Extended the schema to the three-part model (frameworkHints,
knownTraps, shortcuts); added Method E and the native-dialog-trap
category.
draft-han-ai-manifest-00 (2026-04-21)
Initial version.
Author's Address
Won-pyo Han
Individual
Korea, Republic of
Email: pk102h@naver.com
Han Expires 21 May 2027 [Page 41]