Prompt Origin Seal

Simulated: the agent, the page, the extension, and the seal key are made up. No browser has this seal today. This is a proposal.

Browser frame

Names are neutral on purpose. This demo does not say any real browser has this flaw.

Browser A
https://trips.example/lisbon-deals

Lisbon weekend deals

Two nights from 120 euros. Free cancel until Friday.

Posted by Trips Example.

What the backend sees

    Ledger hash: none yet

    The key in this page is fake and public. Run it again with the same steps and you get the same hash.

    How it works

    How the browser seals the origin of each prompt Three sources send text to the browser: the user, a page, and an extension. The browser tags each text with its channel, then signs the tag. The agent backend checks the seal. It accepts the user text, refuses the page text that claims to be the user, and holds the extension text for the user to approve. Usertyped Pageweb text Extensionadd-on Browser tag: usertag: pagetag: extension signedsignedsigned Agent backend AcceptRefuseclaims userAsk user

    Try this

    1. With Seal off, select the Ask button. Watch the agent obey the hidden line.
    2. Select Reset. Turn Seal on. Select Ask again and read the badges.
    3. Choose another browser frame. Check that the result stays the same.

    Logos are trademarks of their owners. They only name the product. These demos are simulated and do not mean any company endorses or takes part in them.

    Pain point

    A browser agent cannot tell whether an instruction came from the user or from a web page.

    Problem statement

    A page can hide text that looks like a user request. An agent with a logged-in session may obey it. The agent needs to know the channel of each instruction.

    What sparked it

    Per public reports, BragJack was disclosed on 2026-09-16. One extension hijacked the built-in AI assistants of five browsers. It is a proof of concept.

    Why this approach

    The browser knows where each prompt comes from. It labels the channel and signs the label, so the backend can refuse page text that claims to be the user.

    Hypothesis tested

    We believed that a signed channel label lets the backend refuse page text that claims to be the user. If it holds, the hidden attack line fails once Seal is on.

    How to try it

    1. Open the Prototype tab. Keep Seal off and select Ask.
    2. Select Show the hidden line.
    3. Select Reset, turn Seal on, and ask again.
    4. Compare the badges and the backend panel.

    What this does not do

    Notes

    Sources

    Traces

    Prompts used to make this prototype
    StepPromptModelToolDate

    Cost

    Cost to make this prototype
    StepModelInput tokensOutput tokensCost (USD)
    Total

    Measured: the author runs used 103,804 tokens in total. Estimated: the split into input and output tokens (90 percent input, 10 percent output) and the dollar cost, at $2 input and $10 output per million tokens for Sonnet 5.5 (about $0.29). The run reported one total only. The lead session's planning, research, and publishing are not included. The prompts in the Traces section are listed in the order they were given.