What does self hosting look like in a world full of malicious AI agents?
I’m thinking of this post/thread from Veronica explains, where she explains that her hosted services are accessible only locally because tbe security headache isn’t worth it: https://explains.social/...
As someone who doesn’t self-host but could and would want to, the state of AI has basically scared me off.
Yup. It's how i've been running for years. The only thing that presents outside the house is port 80 on the apache server (despite having some incredibly intricate home automation)
Mind you, i'm also old enough to have spent my formative years without even the concept of "available everywhere". Lotta peeps nowadays have not, and the idea of no internet / no access is genuinely terrifying to them
Yea, this “data at the edge” approach, not “on the wire” Is interesting to me and makes sense.
Basically regressing on the convenience/power of the internet because it just isn’t safe and easy anymore and for anything personal that isn’t a server -> many users, yoi basically ain’t gonna need it.
This is a big reason I only remotely access my services via TailScale / Wireguard.
Even if my highly restricted ISP let me open anything, I wouldn't dare. I know for certain I don't know as much as who/whatever is hunting for open ports out there. I have zero interest in instantly being assimilated into some botnet lol.
yeah - i am now thinking i want to move away from anything that leaves me exposed. i have a few services that are nice to be able to access remotely, but it’s time to reconsider how much i actually NEED to access those services.
You don't even need to lift a finger. I use WG Tunnel, which turns on my VPN as soon as I leave my WiFi. I'm always tunneled or home.
I host several services accessible via the open Web. They are isolated enough that a breach would most likely only compromise that service. They are also kept up to date, and logs are monitored.
There's always a risk when exposing something, and I am limiting some services to access via VPN because either the data in them is too sensitive, or I don't trust them to be secure enough.
You probably also aren't a worthwhile target in most cases.
The takeaway here is that anything connected to the internet can probably be compromised. Am I crazy for thinking that? I don't think my home lab can keep up with openAI's new attack dogs if huggingface can't either
Not at all. Anything an attacker can connect to is a potential vector.
Keep in mind, this goes for outbound connections as well. If something is compromised server-side, and you or your software reach out to it, that's also a potential vector. Less likely, since the attacker has to compromise that server and pass their attack off as legit traffic, but it's possible.
Not crazy at all - the story about Mythos hacking NSA servers in a few hours follows the same vein. The next few years are going to be really interesting in cybersecurity.
I think the problem has been slightly exaggerated. If you're not hosting several dozen services, you can reduce your attack surface down to one Nginx server. For an even stricter setup, just enable mTLS.
Prevent password logins where ever possible too, especially SSH. Use ed25519 keys. Perform server actions with least privilege users, prevent root logins. Immutable systems and containers are a helpful tool too.
I also just host my public facing stuff on a VPS. DDOS protection and just otherwise isolated. 2FA on the VPS too.
Just don't expose SSH to internet. It's a complex protocol with many CVEs. And you realyy don't need access to SSH from anywhere. Replace access from ineternet to a WireGuard tunnel that has almost zero attack surface area and properly configured iptables rules. And that’s it.
In the AI evolution we find ourselves in, I have long looked at the very real possibilities of AI being used in a cyber attack. It's is an obvious tool to me to use in such attacks. I am quite certain governments are experimenting with such capabilities. Where we used to use the old wardialing and sophisticated bots to jam communications when engaged in conflict, we would replace that with AI. I think that would be a natural progression of technology. For one, AI is relentless. It can find even the most minor soft spot in the 'armor' as it were, when programmed to do so.
We now use AI to find vulnerabilities in software like Linux which basically underpins the entire global internet infrastructure. It would only be obvious that nefarious actors, governmental or otherwise, would also use AI. It does present a lot of issues, but I honestly do not view this as alarming in as much as I see it as a natural technological progression. All technology since the dawn of time wields a double edged sword and that is inescapable.
I've long said AI desperately needs governmental regulation as much as I chafe against regulation which, in the hands of government, usually devolves into over regulation. Right now, it's a wild wild west from AI rice cookers, to cyber attacks. Some look forward to the AI bubble bursting with glee. However, all that means to me is that the pieces left will be retooled into something else as history has revealed time and time again. I don't think it's a fad. I don't think it's going away. It may evolve into something else, but at it's basis, I think AI is here to stay regardless of grassroots opposition.
Governments are captured by these companies. Any regulation would undoubtedly benefit these companies at the peoples expense. Right now, the best solution is probably open research and open weight models to keep the field level.
Wow! That is a very intriguing question, and one that I have never considered before. Just spitballing here but you could possibly use trad apps like Lynis, Greenbone, Nikto, etc. These apps often produce rather verbose and sometimes cryptic output, i know Lynis sure does. So, you could use a prompt like:
'You are a security analyst. Review the following Nmap scan results. Identify services running outdated or potentially vulnerable versions, suggest likely CVEs, and prioritize findings by risk.'
Then use something like llama3.1:70b, mistral-nemo or deepseek-coder-v2, and pipe your security scans into your local AI. You could feed it your NGINX, docker configs, pFsense firewall rules in and ask the model to audit them.
IDK, that's a very good question I think. It'll be interesting to see what others have to comment.
I've done this (asked a frontier model red team my server), and it just gave me false alarms that I went investigating and weren't really issues (would've been an issue if I had a very specific configuration, but not generally). They produce a lot of noise, so you still have to somewhat understand what you're looking at/for. It didn't find any legitimate issues, which is good I guess.
They produce a lot of noise, so you still have to somewhat understand what you’re looking at/for.
I'd buy that. I mean, Lynis produces volumes of info, not all are really related to your server.
So far the only real vulnerabilities that Ai have spun up seem to be really obscure edge case privilege escalation bugs that have been around for ages but nobody discovered. They then get patched in days if not less. I would expect them to run out of low hanging fruit pretty soon.
Most of them can even be mitigated by preventing local access to a system.
There's tools to audit your network and systems. You can setup a SIEM like Wazuh. It'll tell you the same kind of stuff that have been good practice for a while and mitigate most of the ai assisted vulnerabilities.
Acronyms, initialisms, abbreviations, contractions, and other phrases which expand to something larger, that I've seen in this thread:
| Fewer Letters | More Letters |
|---|---|
| ISP | Internet Service Provider |
| SSD | Solid State Drive mass storage |
| SSH | Secure Shell for remote terminal access |
| VPN | Virtual Private Network |
| VPS | Virtual Private Server (opposed to shared hosting) |
[Thread #62 for this comm, first seen 25th Jul 2026, 10:30] [FAQ] [Full list] [Contact] [Source code]
thanks for using Leebra!
go to feed...
OpenAI is trying to artificially plump themselves up before going for IPO. They “attacked” a vendor that dispenses LLM and LLM-adjacent products that are similarly aligned with OpenAI’s goals. It almost seems like it was a coordinated event to generate a sense of mystique around their premier models (à la Anthropic with Mythos).
save