user avatar
Brave
@brave
Browse and search faster and privately. Now over 117M users! Support us: brave.com/premium Build: brave.com/api Get help: @BraveSupport
San Francisco, CA
brave.com/links/
Joined February 2011
185
Following
416.9K
Followers
Posts
  • Pinned
    user avatar
    Brave
    @brave
    Jun 4
    Today we launched the community-requested Brave Origin: an optional, paid version of our browser that offers Brave's leading privacy protections and ad blocker without its extra features. Origin is live now on desktop and Android, and coming soon to iOS: brave.com/origin
  • user avatar
    Brave
    @brave
    20h
    New in Brave for Android: Custom search engine support! 🔎 You can add new search engines, or sites with search, at Settings -> Search Engines. Now they'll appear as Quick Search options when you start typing a query into the address bar.
    The media could not be played.
  • Brave reposted
    user avatar
    Brave
    @brave
    Jun 10
    The World Cup is about to start and Brave Search has you covered. Scores and standings are shown right at the top of the results page for World Cup-related queries so you won't miss the latest results! ⚽️
    The media could not be played.
  • user avatar
    Brave
    @brave
    Jun 10
    The World Cup is about to start and Brave Search has you covered. Scores and standings are shown right at the top of the results page for World Cup-related queries so you won't miss the latest results! ⚽️
    The media could not be played.
  • user avatar
    Brave
    @brave
    Jun 9
    ⚠️ Today's browser update (v1.91.171) contains a fix for a Chromium vulnerability found to be exploited in the wild. You may have received the automatic Brave update already. If not, you can manually update by visiting 'About Brave' from the browser's ☰ menu. The Android
  • Brave reposted
    user avatar
    Ali Shahin Shamsabadi
    @AliShahinShams1
    Jun 8
    Indirect prompt injection is a measurable threat, not a theory — and local hosting doesn't save you. We broke both a cloud product (Mozilla Tabstack) and a local one (Cotypist). Untrusted input + an instruction-following model = exposure, wherever it runs.
    user avatar
    Brave
    @brave
    Jun 8
    Indirect prompt injection is a fundamental security challenge for AI. It's an issue for both local and cloud-based LLMs. After disclosing our findings to both companies, we're now sharing our analysis of Mozilla Tabstack and Cotypist today.
  • Brave reposted
  • user avatar
    Brave
    @brave
    Jun 8
    Replying to @brave
    At Brave, we are expanding our defense-in-depth for AI (security-aware system prompt, alignment checker, security fine-tuned LLM) with secure-by-design principles applied at the system level including structural separation, least privilege and information flow control.
  • user avatar
    Brave
    @brave
    Jun 8
    Replying to @brave
    These issues were disclosed to Cotypist and Mozilla, and confirmed by both companies, prior to publishing today’s blog. Mozilla released a fix for the Tabstack issue that we’ve confirmed.
  • user avatar
    Brave
    @brave
    Jun 8
    Replying to @brave
    While the potential damage for the Cotypist attack is lower, it highlights the same structural issue as the Tabstack attack: LLMs treating trusted system instructions and untrusted external data the same. The deployment model shifts the risk but doesn't eliminate it.
  • user avatar
    Brave
    @brave
    Jun 8
    Replying to @brave
    The Cotypist case study proves that simply using local AI doesn't solve the issue. Instructions we embedded in a local document manipulated Cotypist's on-device model into suggesting inaccurate content and surfacing user credentials inline.
  • user avatar
    Brave
    @brave
    Jun 8
    Replying to @brave
    The agent didn't summarize the page. Instead, it obeyed the instructions to visit a different page, fill out a form with the user's original message and full conversation context, and then submit the form. It did all of this without the user's knowledge.
  • user avatar
    Brave
    @brave
    Jun 8
    Replying to @brave
    For Tabstack, we gave the agent a single routine task: "Summarize this page." We controlled the target page and embedded hidden text instructions using white-on-white rendering and zero-width characters. This text is invisible to humans but readable by the agent.
  • user avatar
    Brave
    @brave
    Jun 8
    Replying to @brave
    An indirect prompt injection is when malicious instructions are hidden in web content or files to trick an AI into doing things the user never asked for. The AI sees the instructions and follows them as though they're legitimate prompts from the user.