Brave shows invisible instructions hijacking AI browsers

Why the screen matters21 October 2025

Brave’s researchers showed that text almost invisible to people, in images or on pages, was read by AI browsers as instructions from the user, letting a web page steer an agent acting with the user’s logged-in privileges.

Position
Content on the page, read by an AI agent

What happened

On 21 October 2025 Brave published research on AI browsers. In Perplexity's Comet, faint text hidden in an image, such as light blue on yellow, was extracted from a user's screenshot and passed to the model as if it were part of the user's request. In Fellou, simply navigating to an attacker's page sent its text to the model, where it could override what the user asked. Brave reported the issues to the vendors before publishing. Its conclusion was that AI browsers fail to separate trusted user input from untrusted web content while acting with the user's logged-in privileges.

The ATG nexus

The reader of the screen is no longer only a person. An agent reads the page too, and treats what it reads as instructions it may act on with the user's authority. That turns content into code, which the research calls Data as Code, and shortens the distance from showing something false to doing something harmful.

Limits of the comparison

These were demonstrations reported to vendors, not documented attacks on users.

On this site

Sources

  1. Unseeable prompt injections in screenshots: more vulnerabilities in Comet and other AI browsersBrave, 21 October 2025Primary source
  2. Agentic Browser Security: Indirect Prompt Injection in Perplexity CometBrave, August 2025Primary source
  3. Unseeable prompt injections (link post)Simon Willison's Weblog, 21 October 2025

Back to this event on the timeline