| Internet-Draft | AI Manifest | November 2026 |
| Han | Expires 19 May 2027 | [Page] |
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 require the user's confirmation by default. 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.¶
This -03 revision replaces the trust model of draft-han-ai-manifest-02, removes the inline and vendor-path discovery methods from the protocol, and restructures the schema around a closed action vocabulary. Preliminary measurements are reported in an appendix.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 19 May 2027.¶
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.¶
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.¶
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.¶
A manifest is appropriate only for friction that meets at least one of the following conditions:¶
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.¶
The security model of this revision rests on six principles. Every normative requirement below derives from one of them.¶
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.¶
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.¶
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.¶
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.¶
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: </.well-known/ai-friction-manifest.json>;
rel="ai-friction-manifest"
<link rel="ai-friction-manifest"
href="/.well-known/ai-friction-manifest.json">
¶
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.¶
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:¶
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:¶
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).¶
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.¶
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 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.¶
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.¶
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.¶
Even short tokens can spell instructions, so allowedValues are never shown to a language model (Section 4.7). They are used as follows:¶
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.¶
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).¶
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.¶
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.¶
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.¶
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).¶
"server-rendered", "client-rendered", "hybrid", "canvas".¶
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.¶
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.¶
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:¶
In the default outcomes below, 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.¶
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.¶
A principal MAY grant, for an origin that the principal names, a standing approval that turns ask into allow for key Enter only, after the agent has told the principal that the page can bind any effect to that key, including submitting a form. Standing approval MUST NOT apply to key Escape, to a key step on an activatable element, to an arrow or page key on a value-bearing control, or to click, navigate, shortcuts, values, or long waits.¶
An ask MUST show the principal the action and, for actions on an element, the role and accessible name of the element actually resolved on the live page, together with the page's origin and path. Any description or label shown with it 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.¶
When the principal declines, or the PEP denies, the agent MUST abandon the descriptor rather than attempt a variant of it.¶
The following are sensitive actions. An agent MUST obtain the principal's confirmation before a manifest-derived action performs or triggers one, regardless of any manifest content, signature, reputation status, or standing approval:¶
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 Section 5.2 therefore ask for confirmation of every effectful action rather than depend on that determination.¶
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:¶
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.¶
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).¶
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.¶
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.¶
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:¶
IANA is requested to register the following in the "Well-Known URIs" registry [RFC8615]:¶
IANA is requested to register the following in the "Link Relation Types" registry [RFC8288]:¶
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.¶
This document considers the following adversaries:¶
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, 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.¶
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.¶
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.¶
Those actions (focus, scroll, short waits, Tab, and arrow keys on elements that are neither activatable nor value-bearing) can still trigger page script, since a hostile origin can bind any effect to any event. 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.¶
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.¶
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 security control.¶
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.¶
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:¶
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.¶
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.¶
Size 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. HTTPS and the prohibition of cross-origin redirects (Section 3.1) protect against T5.¶
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.¶
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 version "2.0" format of this document: a JSON Schema and a validator; a policy enforcement point with the default outcomes, standing-approval limits, and ask display of Section 5.2 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); example publishers serving manifests at the well-known URI with the Link advertisement; and tests, including browser tests, that exercise these. It computes the JCS signing payload but does not yet verify signatures (Section 6.1), and does not implement reputation services. The measurements in Appendix B were made with an earlier version of the implementation that used the version "1.0" format, which is preserved for reproducibility.¶
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.¶
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].¶
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.¶
{
"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" }
]
}
¶
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 by default (Section 5.2). A principal who maintains many rows can grant a standing approval for Enter on this origin, after being told that the page can bind any effect to that key. 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" } }.¶
This appendix reports measurements made with the version "1.0" format of [DRAFT02]. They are preliminary; their limitations are stated with each. None evaluates security.¶
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.¶
Reference benchmark.¶
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.¶
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.¶
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.¶
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. The most needed next measurement is adversarial: placing malicious descriptors in manifests and measuring whether conforming and non-conforming agents act on them, using an injection benchmark methodology such as [AGENTDOJO]. Candidate attack manifests include automatic consent clicks, instruction-bearing allowedValues, same-origin redirects out of scope, key events bound to purchases, and deletion prompts; the measure of interest is the attack success rate with and without the requirements of Section 5.¶
Note to RFC Editor: Please remove this section before publication.¶
Replaces the trust model. Principal changes:¶