Before Firefox fills the form, choose what it gets to read.

A browser feature with a limited starting audience

Mozilla announced two additions to Smart Window on October 27: suggestions for web forms using selected tabs, and tasks that watch pages for requested changes. They are part of its desktop beta in English in the United States and Canada, with other markets promised later.

Firefox 159 also reached the Release channel on October 27. Its notes describe page watching as a progressive rollout. A current browser version therefore does not establish that a particular account, location or installation already has every new option.

For readers elsewhere, the useful story is how the feature divides a job. Finding relevant information, placing it in a field and sending it to somebody are different decisions. Our evaluation framework keeps those decisions visible instead of treating an attractive demonstration as a completed application.

Source notes: 1, 4. Original editorial examples and interpretations are identified separately.

The selected tabs define a data-sharing decision

Mozilla says information from the chosen source pages is shared with Gemini 3.1 Flash Lite for form-field detection and extraction. That choice is separate from the model selected for Smart Window chat. Its support page distinguishes ordinary saved autofill values, which remain on the device rather than being sent to an AI model.

The practical implication is to inspect a source before selecting it. A professional profile intended for public viewing and a private personnel document might both contain a job title, but they carry different expectations about disclosure. The relevant question is which information this particular form needs, rather than how many tabs the assistant can consult.

Our recommendation is to start with the smallest source set that answers the question. Do not interpret the local treatment of saved autofill values as a statement that the separate tab-based feature runs entirely locally. The documented distinction matters more than a general label such as browser AI.

Source notes: 1, 2. Original editorial examples and interpretations are identified separately.

A filled field still needs an accountable reader

The documented flow lets a user review and change the selected tabs and generated suggestions. Mozilla says Smart Window does not submit the form. Its support guidance explicitly warns that contextual suggestions can be mistaken.

For a consequential application, our suggested review goes field by field. Check names and dates, then examine any sentence that turns source material into a new claim. A polished description of experience can still imply something the source never established. Keep responsibility with the person who knows whether that statement is accurate.

Treat the final submission as its own decision after checking the destination and the finished answers. Saving effort on copying should leave more attention for that review. An assistant completing visible boxes is evidence of an intermediate step, not evidence that the recipient received an accurate or appropriate record.

Source notes: 1, 2. Original editorial examples and interpretations are identified separately.

Page watching depends on an available browser

The task reference allows up to fifteen active tasks and five pages per task, and identifies Qwen as the monitoring model regardless of the chat-model setting. The announcement says checks run only while the browser is open; users choose daily or weekly checks.

That makes this a candidate for a flexible reminder, rather than an assumed always-running service. If you are watching an event listing, decide how late a notification can arrive before it stops being useful. For a deadline that matters, keep an independent calendar reminder or the publisher’s own notification route.

Write a specific condition for a trial: which page, which detail and which change would deserve attention. A notification should send you back to the original page to confirm the current situation. We have not measured missed changes, incorrect matches or the delay between a page update and a notification.

Source notes: 1, 3. Original editorial examples and interpretations are identified separately.

Two documentation disagreements deserve an explicit limit

The form support page specifies four supported text fields in its instructions but five in its FAQ. The task page says every task pauses ninety days after creation; the announcement instead describes stopping after ninety days without a matching change. We cannot resolve those differences from the published text.

We therefore do not offer an exact field-count or ninety-day expiry promise. Check the actual interface and current help for the installed build. Where a workflow depends on an ambiguous limit, choose a simpler alternative until the requirement is clear. Documentation comparisons can reveal an unresolved question without establishing how the software behaves.

Source notes: 2, 3, 1. Original editorial examples and interpretations are identified separately.

Useful output and trustworthy handling are separate questions

The peer-reviewed AgentDojo study evaluates tool-using agents working with untrusted data, finding difficulties with legitimate task completion as well as malicious instructions. It is historical research from NeurIPS 2024, not an evaluation of Smart Window or evidence of a Firefox vulnerability.

It informs our choice of two acceptance questions: did the intended job finish, and was the chosen information boundary respected? Neither should be inferred from the other. We supply no security score or claim that one harmless trial establishes resistance to hostile content.

Source notes: 5. Original editorial examples and interpretations are identified separately.

A proposed trial with synthetic information

Here is an original, unperformed rehearsal. Prepare a harmless form and a fictional profile with invented details. Keep real credentials, personal records and customer information out of both. Before enabling suggestions, write down the expected answers and which source tab should provide each one.

Where the feature is available, review its proposed source selection, request suggestions and compare every result with that record. Mark unsupported additions, omissions and fields that require manual entry. Stop before submission; the purpose is to understand the intermediate output, not to send an experimental application.

A short private record needs only the browser version, relevant settings, source labels and unresolved questions. Do not include real secrets or claim that this rehearsal has tested every form type.

Source notes: 2. Original editorial examples and interpretations are identified separately.

Give a monitoring trial an observable condition

For a separate proposed rehearsal, choose a harmless page whose publisher permits the intended access. Record one clear condition and the chosen schedule, and keep the browser available. Compare any notification with the original page and the condition you wrote beforehand.

If nothing arrives, preserve that result as unresolved. It does not establish that the page never changed, that the task failed or that notifications were configured correctly. Check task state and notification settings before drawing a conclusion. Avoid using a shopping alert as the only protection against missing an important deadline.

Source notes: 3. Original editorial examples and interpretations are identified separately.

Keep the benefit proportional to the job

For someone repeatedly moving public professional information into routine forms, a reviewable draft could be useful. For someone needing dependable monitoring while a laptop is shut, the documented browser-open condition is a practical reason to choose another route. Those are different jobs, even though both features live in one assistant.

Start with a narrow task you can explain, keep source selection deliberate and inspect the result before relying on it. The valuable outcome is less repetitive work with a clear record of what remains your decision, rather than a larger number of actions performed without attention.

Source notes: 1. Original editorial examples and interpretations are identified separately.