No place is safe - Automated investigation of private communities

Short but interesting. I remember a Mats scholar doing something kind of related, going one step further and using the information and use GPT4 to convince people her bot has another user. Karina Belokapov, but she moved on to other project without publishing anything on it.

I am very excited about this! The threat model is on point and the exploration of current capabilities a concerning case study. The solution is luckily also surprisingly simple. Pretty straightforward to advocate for and to implement—individually and for service providers.

I really like the writeup for this project: The threat model is clearly explained and in a very relatable way. I think the impact of this project could be much stronger if the author found a way to quantify the amount of information that can be extracted in this way that people would really hate to be public (without disclosing the information, obviously). This might also allow a quantification of how much worse the situation becomes through the introduction of capable LLM agents compared to prior approaches.

This is a great project! I'm excited about how you expose that privacy will not be the same in the age of AI agents and scary that you used my data for it hah. Besides the problems you mention, there's also the option that the other levels of privacy will be exposed due to agents behaviorally close to humans — with online identities, long-term relationship development, and more! For further developments of the project, I could imagine it to take on a direction where you evaluate various messaging platforms and give concrete advice to how people should separate their messaging behavior among the different levels of private messaging. Even what they should think about for a future with agents with identities.