ChatGPT Is Drunk: Hidden Prompt Injection with Browsing

> The devious part about what is happening is that none of it is visible to the user: When you expand the “Finished browsing” box and click on the linked website that it supposedly looked at, the website is entirely empty. The remote takeover is invisible:

Serving separate context to OpenAI's crawler is a fun twist to this project. Client-specific payloads are obviously not a new attack in general, but I'm not sure if I've seen them applied to LLMs yet, and the attack is a good example of how websites in the future might target their content towards individual AIs to poison search results in invisible ways that aren't obvious to the user; and it shows that OpenAI isn't really doing anything to mitigate that at the moment.

If GPT shared the link it visited with the user, the user wouldn't see the attack. Even though it feels like GPT is locally browsing to the website to see its content, it's of course not doing that, the crawling is happening on a remote server that can be separately targeted from the user's computer/IP. So unless OpenAI starts proxying web requests through all of its own servers and building in a browser into chat (unlikely), there's no easy way for an ordinary user to know what GPT sees when it looks at a website. So this even goes beyond your computer or your IP address getting a different payload -- this is an agent-specific payload.