TL;DR: it looks like there are routers on the market that use gpt APIs to classify new, never seen before websites and add them to the blocklist in real time
This allows everyone to run something like the “great firewall of China” - and imagine what a government could do
Full story:
At work the boss got persuaded to pay some thousands for a fortinet firewall to cripple the customers free wifi (the extremely stupid idea is to sell them an unblock code, but I live in a country where with 10 euro per month people can get 100gb of 5G connection, who’s going to pay?)
I tried that network and I was really shocked how crippled it was. Boss decided to block anything related to gaming, for example. You visit a small game developer page and it initially works but after a few minutes, you get a “blocked” page (but customers can’t see that because nowadays everything uses HTTPS and they don’t have the self signed CA on their system - they just see HTTPS certificate error). I tried multiple times but always the same result, after a few minutes is blocked.
Everything that corporate thinks it’s not appropriate, it’s blocked!
I felt more frustrated using this network than the time that I lived in China! (Left a few years before COVID, don’t know the internet situation now)
When I came back home I took a shower and I thought to it under the hot water. At home I’m using gpt4o in karakeep to classify my bookmarks… and a router can also do the same.
Boss decided to block anything related to gaming, for example. You visit a small game developer page and it initially works but after a few minutes, you get a “blocked” page (but customers can’t see that because nowadays everything uses HTTPS and they don’t have the self signed CA on their system - they just see HTTPS certificate error). I tried multiple times but always the same result, after a few minutes is blocked.
That’s not “AI”. That’s just a “man in the middle (mitm)” attack. AKA, https proxy. Fortinet firewalls have been capable of that for many years. It’s not uncommon for businesses (e.g. banking and finance) to proxy all internal web traffic in order to make sure personally identifiable information isn’t being shared with shady websites.
Proxying traffic on the guest WiFi is pretty sketchy though.
I think I lost you somewhere; how do LLMs factor into this?
They may or may not be used here. You could use LLMs to parse the content of sites being visited by web clients on your network. Then, ask the LLM whether the content includes certain topics or is work related. Based on the results of that, you add/remove the site from a blacklist.
Is this better than just string matching? I would say likely so, though more stochastic in the results. It would let the LLM provide summaries/context of the pages, and not by just confined to specific strings in a list. It might be better ramble to handle context and complexity of the desired outcomes.
For example, there was a paleontology conference at a hotel once that was stuck behind a firewall blacklisting all sites with the string ‘bone’ in them. Completely ridiculous. The string ‘bone’ has different meanings based upon context, which simple string matching cannot provide, but an LLM might be better and identifying and acting accordingly.
Cool…I worked in this school once that blocked any website about calculus because of the “cu” part of the string that means ass hole in my language. It was funny.
What about encrypted DNS?
It’s all blocked and I didn’t find any single VPN to work (didn’t try those with shadowsocks)
Tailscale worked but only if i login using 5G, then after choosing my exit node on my own server, connect to the crippled WiFi
Modern firewalls are very good at sniffing out what a host is doing.
