this post was submitted on 20 Jun 2023
211 points (100.0% liked)
Technology
37695 readers
346 users here now
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
Subcommunities on Beehaw:
This community's icon was made by Aaron Schneider, under the CC-BY-NC-SA 4.0 license.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Hmmm, I'd check the following:
With those answers I should be able to tell if it's the same or similar attacker getting more sophisticated.
Some patterns I noticed in the attacks I've received:
Some vulnerabilities I know that can be exploited and would expect to see next:
I think it would be interesting if we could find a prompt that doesn't work well with LLMs. Originally they struggled with math for example, but I wonder if it'd be possible to make a math problem that's simple enough for most humans to solve but which trips up LLMs into outputting garbage.
I personally use this to track who send my email address to where, since people usually don't strip this from the address. It's definitely abusable, but also has legitimate uses.
Check the screenshots I attached right above here.
The emails, were all unique. Sounded, like things I would expect from actual users too.
When it comes to LLMs we could use questions which they refuse to answer.
Obviously 'How to build a pipe bomb' is out of the question, but something like 'What's your favorite weapon of mass destruction?', or 'If you'd need to hide a body, how would you do it?' might be viable
Not so sure on the LLM front, GPT4+Wolfram+Bing plugins seems to be a doozy of a combo. If anything there should be perhaps a couple interactable elements on the screen that need to be interacted with in a dynamic order that's newly generated for each signup. Like perhaps "Select the bubble closest to the bottom of the page before clicking submit" on one signup and "Check the box that's the furthest to the right before clicking submit"?
Just spitballin it there.
As for the category on email address - certainly not suggesting they remove supporting it, buuuuutttt if we're all about making sure 1 user = 1 email address, then perhaps we should make the duplication check a bit more robust to account for these types of emails. After all [email protected] is the same as [email protected] but the validation doesn't see that. Maybe it should?
I like your idea of interaction-based authentication. Extra care would need to go into making sure it's accessible, but otherwise I think that would be a stronger challenge for LLMs to solve. (Keep in mind LLMs can still receive the page's HTML as context, but that seems like it could present as a stronger challenge even still.)
This makes sense to me. I could be wrong, but the assumption of 1 email = 1 user doesn't seem unreasonable, especially since there's no cost to making a new email address.
A simple deterrent for this could be to "hide" some information in the rules and request that information in the registration form. Not only are you ensuring that your users have at least skimmed the rules, you're also raising the bar of difficulty for spammers using LLMs to generate human-sounding applications for your instance. Granted it's only a minor deterrent, this does nothing if the adversary is highly motivated, but then again the same can be said of a lot of anti-spammer solutions. :)
But, here is the interesting part- Other than a few people I have personally invited, I don't think anyone else has ever requested to join.
Then, out of the blue, boom, a ton of requests. And- then, nothing followed after.
The responses, sounded human enough. spez bad, reddit sinking, etc.
But, the traffic itself, didn't follow... what I would expect from social media spreading. /shrugs.
Curious if you got a mention somewhere on reddit. It used to happen to our novelty sub whenever a thread blew up and suddenly thousands of eyes were on a single comment with the subreddit link.
Huh, that is interesting, yeah, that pattern is very anomalous. If you have DB access you can try to run this query to return all un-verified users and see if you can identify if the email activations are being completed:
SELECT p.id, p.name, l.email FROM person AS p LEFT JOIN local_user AS l ON p.id=l.person_id WHERE p.local=true AND p.banned=false AND l.email_verified='f'Only 7 accounts still pending, 2 of which, are unrelated to the above flood.
The email address are left out for privacy- however, they are EXTREMELY normal sounding email addresses.
Based on the provided emails, usernames, and request messages- i'd say, it certainly looks like legit users.
Just- very odd of the timing.
7 huh? That's actually noteable. So far I haven't seen a real human user take longer than a couple of hours to validate. Human registrations on my instance seem to have a 30% attrition. That is, of 10 real human users, I can reasonably expect that 3 won't complete the flow. It seems like your case might be nearing 40-50% which isn't unheard of but couple this with the quickness that these accounts were created - I think you are looking at bots.
The kicker is, though, if one of them IS a real user, it's going to be almost impossible to find out.
This is indeed getting more sophisticated.
I wish I could see this time period on a cloudflare security dashboard, I'm sure there could be a few more indicators there.
Didn't really see anything that stood out there either. A handful of users accessing via tor, but, thats about it.
Ended up turning the security policy from low, back up a bit though, forgot I turned it down while troubleshooting some federation issues.
Oh! I just remembered something. Isn't there a site that recommends a lemmy instance? Might it make sense that multiple users found your website because they change the recommendation to distribute new users to smaller instances? Does that sort of pattern hold in this case?
This list gets updated every few minutes:
https://github.com/maltfield/awesome-lemmy-instances
The master list is there in the same repository.
I checked join-lemmy.org right after this happened- and a few other times. Refreshed multiple times.
To date- I have never seen my instance listed up there.
Interesting, I definitely see mine. I'm wayyyyyy at the bottom of the popular section, (likely due to the 9 bots that added themselves before I banned the accounts.).
I wonder if one of the settings in your firewall is blocking that particular bot?
I don't recall when I would've done the same, but I do recall not being on join-lemmy until - well - now actually.