Security Alert: AI-Powered Browsers at Risk of Indirect Prompt Injection
Security researchers at Brave Browser have uncovered a critical vulnerability affecting AI-powered web browsers: attackers can embed hidden instructions in webpages that the browser’s AI agent interprets and executes as if the user had given them.
What the Flaw Is
When a browser with AI features is asked to summarise or navigate a webpage, it often passes the page’s entire content to its AI model—including maliciously hidden instructions. Because the AI cannot reliably distinguish between user commands and hostile content, it may obey those hidden prompts.
Why It Matters
These AI agents usually operate under the full privileges of a logged-in user — meaning the browser could access email, banking, cloud accounts or corporate systems. Traditional web-security safeguards such as same-origin policy become ineffective when the AI acts with the user’s credentials.
What You Can Do
- Avoid using AI-agent browsers for sensitive tasks (e.g., banking, email) until they mature.
- Use separate browsers for general browsing vs. authenticated sessions.
- Ensure any agentic feature asks for explicit user consent before taking cross-domain actions.
- Keep whichever browser you use updated and monitor vendor security advisories.