Yeah, it happens: https://blog.torproject.org/bad-exit-relays-may-june-2020 https://blog.torproject.org/bad-exit-relays-may-june-2020 One reason why it's not d

Yeah, it happens: https://blog.torproject.org/bad-exit-relays-may-june-2020 https://blog.torproject.org/bad-exit-relays-may-june-2020 One reason why it's not devastating to the network as a whole is that the process for getting your relays to make up such a large fraction of the network is social. If you run a ton of capacity, especially if added all at once, people are going to notice, and reach out to find out who you are (and if they can't, expect to get removed). This means that while yes, you can do this (and as above, people have), once it's detected, all of your resources are dropped at once, and you have to start a pretty expensive and time-consuming process over again. It's also the case that adversaries generally don't collude, so e.g., the above attack was for cryptocurrency theft, and those adversaries likely aren't working with the FBI or China to deanonymize circuits. This means you only have to worry about a few of these happening at a time, which makes it easier to detect (pull on one thread, and the rest start to unravel). That said, just based on the blog post above, it's something that TPO seems to be thinking about new ways to address, and sibyl detection has a long history of research in the academic community as well that has plenty of space left to explore. Something like Salmon[0] is in the process of being implemented by TPO for bridge distribution[1], and the constraints for this reputation problem are far less onerous than that setting. [0] https://content.sciendo.com/view/journals/popets/2016/4/article-p4.xml?language=en https://content.sciendo.com/view/journals/popets/2016/4/arti... [1] https://gitlab.torproject.org/tpo/anti-censorship/bridgedb/-/issues/31873 https://gitlab.torproject.org/tpo/anti-censorship/bridgedb/-...