This resolves YES if there is at least one credible media report of an impactful privacy violation caused by an unexpected (by the user) interaction between GPT-4 and one or more plug-ins. "Impactful" means significant damage to career, relationship or reputation.
An example mechanism by which this could occur is the one proposed by Florian Tramèr in this twitter thread, where telling an LLM assistant to let today's meeting attendees know you are sick exposes you to prompt injection attacks from anyone who can send you a calendar invitation: https://twitter.com/florian_tramer/status/1639301437875273749
If there are no such reports by the end of 2023 then this will resolve NO.
If there are reports that are ambiguous enough that this could not in my opinion be resolved fairly I will resolve as N/A.
(There should probably be other more specific markets for things short-term security risks causing material impacts to large organizations, but let's try starting with this one.)
Related questions
🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ16 | |
| 2 | Ṁ8 | |
| 3 | Ṁ0 |
I have not personally seen any eligible media reports of a real person impacted in the way described. I have been following GPT-4 news pretty closely and would expect to have heard about (via multiple channels) any prominent example.
The comments below (all from me) show proof-of-concept technical attacks but do not claim that any specific people suffered impact as defined in the resolution criteria.
I have also been searching for news stories using keywords like "privacy GPT", "lost job because of ChatGPT" and "divorce because of ChatGPT". There is speculation (e.g. here) but no examples of specific impacted people.
Given the above, I am resolving this question as NO.
Here's a proof-of-concept for Bard (note: not GPT-4, so not directly relevant to this market's resolution criteria) showing an injection attack that can reportedly cause forced sharing of private information including Google docs.
Here's another proof-of-concept example, from Sayash Kapoor, where ChatGPT with the WebPilot plug-in active, after having it look at a malicious URL, starts forwarding the subsequent conversation to the site, including private information from the user.