review prompt behavior of the page content the smart actions toolbar prompt sends before enabling it
Categories
(Core :: Machine Learning: Frontend, task)
Tracking
()
People
(Reporter: Mardak, Unassigned)
References
(Blocks 1 open bug)
Details
(Whiteboard: [aiasst])
Bug 2091509 adds browser.smartwindow.actionsToolbar.prompt for internal Nightly testing. Its default template includes {page_text} (up to 10000 characters) and {page_anchors}, which generateFromPrompt in SmartActionsToolbar.sys.mjs reads with PageExtractor and inserts into the prompt as is. Only {title} and {open_tabs} go through sanitizeUntrustedContent. Sidebar conversation starters send only the page title and URL, so with this prompt every visited page sends its content to the model. Raised in review of D342411.
The model output becomes toolbar buttons that run on click. Chat and search actions run exactly their label, so what runs is what the button shows. A jump action runs a generated fragment that the button does not show, but it only scrolls the current page.
Before the toolbar pref can be enabled for anyone beyond the internal team, get a review from the AI security reviewers covering:
- how page text and anchors are marked as untrusted in the prompt, as other Smart Window features do for page content
- which context the default prompt should send, and whether page content should be read on every visit
- limits on generated labels and details, including the length of a search query
- whether the remaining prompt injection risk in what the buttons show and do is acceptable
firefox.js carries a comment on the pref pointing at this bug.
Updated•2 days ago
|
Description
•