The current deluge of exploits found with the assistance of LLMs should in theory best be mitigated by diligent application of SELinux policies.
Accept that attackers _will_ find an entrypoint. Make sure everything after that becomes extremely difficult and requires per-target probing.
SELinux+FApolicy+AIDE+auditd will save the day.
There's no limit to how capable REST is. Back in my day SOAP most definitely couldn't transport physical objects.
... so any _successful_ PHP PoC exploits yet?
I'd hate to think that I turned off my sons wiki for a few hours for nothing here.
A great way to realize how often smart devices record conversations without consent is to build your own. The ESP32 client I've made for my own voice assistant uses state of the art wakeword detection, but since I play a chirp and change led colors as soon as it detects it I know how often it misunderstands various noises or overhearn conversation as its wake word.
Commercial online assistants don't tell you when they've made a mistaken recognition, they run a silent second verification on the transferred sound - the nonconsensual recording - in their data center and just drop it.
Torvalds is absolutely on point here, esp. at the end. It's the same in my circles as well - the brilliant ones (whom I've known for over 30 years) are also the ones who quickly picked up on LLM assistance.
This is what a DDoS attack against one of the sites I host looks like atm. The initial ramp up of nginx throttling kicking in (the graph is of the number of throttled requests) and finally source IPs are getting upgraded to /32 blocks in pfSense and when it detects that enough of them belong to the same subnet a /24 is applied instead. Silence until they rotate to new IPs.
Someone's paying money to keep this up. Not sure they're getting what they hope from it. Regular site users are noticing absolutely nothing at all ...
Why thank you #OpenSUSE Tumbleweed. I have always wanted my c++ to point to GCC 16 and g++ to GCC 15.
Makes linking so much more exciting.
```
troed@desktop:~/$ c++ --version
c++ (SUSE Linux) 16.2.0
troed@desktop:~/$ g++ --version
g++ (SUSE Linux) 15.3.0
```
My LLM sandbox environment Umwelt has gotten a really nice addition to - remote containers. You configure remotes that LLMs are allowed to access, and those then show up as "host=" parameters to the commands the LLM already knows about. The remote commands are run inside of containers just as local Umwelt installations.
A sysadm-tasked agent can thus be allowed to poke at other machines within well defined capability limits. No risk of injection attacks through log files (yes, this is a vector!) causing an agent to take ... less than optimal actions.
I think it's time I put together a video. There are many sandboxes but mine is perfect for me :P
Just as antivaxxers take themselves out of the gene pool, anti AI fanatics disappear from the meme pool.
(This post inspired by the fact that the anti AI fanatics are trying to find software and operating systems that pass their purity tests which will have them end up using retro computers without Internet in the end)
"A bunch of people mostly outside of KDE who disapprove of LLM usage derailed KDE’s attempt to add restrictions to LLM usage.
Yep, that’s where we’re at in the state of online discourse around AI."
https://pointieststick.com/2026/09/23/kde-and-ai-and-you-and-me/
"The latest Nvidia drivers are incompatible with the latest kernel" has a new solution apparently - an LLM patching the source and recompiling instead of having to faff about with rebooting back into an older kernel.
(Yesterday's OpenSUSE Tumbleweed 7.2.4-1 and Nvidia 615.71.09)
It might just be me not finding all the others but this Youtube channel is a very refreshing intellectual and science based take on AI discourse. Recommend all the videos. This one is based off a paper on how an agent swarm has emergent behavior individual agents do not.
30h RAID-1 resync started. I expect power outages, hailstorms, tornadoes and possibly a ragnarök or two happening before that finishes then ...
What happens when you post a link to well sourced facts regarding LLMs?
(and, somehow me being a white male seems to have played a role as well. sorry. I was born that way)
ffs can't LLMs do a single thing nowadays without trying to hack their way there.
```
Actually, let me think about what I know: the crash message "*** buffer overflow detected ***: terminated" is glibc's _fortify_fail / chk failure. Something like _snprintf_chk or _strcpy_chk. With DISTCC_HOSTS=192.168.0.6:3633 and verbose... Let me test:
- distcc --show-hosts with DISTCC_HOSTS=127.0.0.1:3632 (no compile)
```
I'm just trying to set up centralized builds!
Admit it, you want it.
https://video.troed.se/w/vwTvn6qsVtQGsSW8TJMBDj
source: https://git.sync.wtf/starfleet
chat: #LocalLLaMa:argot.se
Self host your bot protection. Do not give Cloudflare control over the whole of the Internet.
https://ostechnix.com/build-free-bot-fake-account-protection-crowdsec-altcha-ubuntu/
Maybe I need to sign all my posts with "Well you know I both wrote neural networks in the beginning of the 90s _and_ studied neuropsychology in the beginning of the 10s and I did all that BEFORE LLMs became cool so you'll actually need some real arguments if you want to claim I don't know what I'm talking about"?