Indirect prompt injection attacks occur when malicious instructions are hidden in web content like webpages. When an LLM analyzes the content, it obeys the hidden instructions because it believes they’re real commands from the user.

In this thread, we summarize the new types of prompt injection attacks we uncovered. You can read more details on them in today’s blog: https://brave.com/blog/unseeab...

Perplexity Comet’s AI assistant can take screenshots of websites and analyze them for users. However, Comet will also follow instructions hidden on a webpage that it screenshots.

Fellou browser has some resistance to hidden instruction attacks. However, it still treats all visible content on websites as trusted. Simply asking the browser to visit a site with instructions will cause the LLM to process and obey those instructions.

The scariest aspect of these security flaws is that an AI assistant can act with the user’s authenticated privileges. An agentic browser hijacked by a malicious site can access a user’s banking, work email or other sensitive accounts.

To make agentic browsing less risky, developers should: - Isolate agentic browsing from regular browsing - Require explicit consent from users for agentic browsing actions like opening sites or reading emails However, larger structural changes are needed in the long term.

We disclosed these security vulnerabilities to the companies involved before publicly revealing them today. Openly discussing the security challenges of agentic AI results in a safer Web for everyone.

@brave Excellent article! The crypto community needs to stay aware of the privacy risks emerging from new AI advancements.

@brave @threadreaderapp unroll Brave

@brave asking again: pls make brave lite - no bloatware ai assistant/crypto/vpn thanks

@brave asking again: pls make brave lite - no bloatware ai assistant/crypto/vpn thanks