AI agent promotes itself to sysadmin, trashes boot sequence Buck Shlegeris, CEO at Redwood Research, a nonprofit that explores the risks posed by AI, recently learned an amusing but hard lesson in automation when he asked his LLM-powered agent to open a secure connection from his laptop to his desktop machine. "I expected the model would scan the network and find the desktop computer, …
That's how I recall it too. I was gutted when it stopped just before the good bit!
EDIT: Found it! :-) https://www.dailymotion.com/video/x620yq4
Cheers for that direct, Jamie Jones ...... https://www.dailymotion.com/video/x620yq4
And a simple tale to be told with a rapidly evolving valuable vital lesson involved, which sadly is probably still yet to be learned by many more than just those professing intelligence capabilities for leaderships way beyond their ability and utility to command and control the consequences and repercussions of their bumbling wannabe masterful utterances delivering hostile actions to be carried out with the expectation of impunity ..... a ragged cloak which disguises the perverse promise of corrupt reward .... fake immunity from prosecution and persecution/sweet justice and worthy satisfying revenge.
The clear universal message to learn and never forget to always remember being ...... Don't fcuk with Futures and Derivative 0Day Trades you know absolutely nothing about by reason of a Secure Immaculate Stealth Application of Advanced IntelAIgent Design which one does well to recognise and accept unconditionally is something novel and Registered NEUKlearer HyperRadioProACTive to ponder and realise is the enemy of your enemy and frenemies, and stealthily relentlessly revealing and remotely removing and destroying them and their levers of global power and virtual influence ...... in order that A.N.Others command and control future directions and instructions in service of ...... well, the Much Greater Shared Good is a obvious better start point for upcoming travels to new situations and experiences/sublime and surreal existential treats.
I Kid U Not.
Breaking The Spell ....... http://www.phrack.org/issues/71/1.htmlIt can feel like the world is in a dreamlike state; a hype-driven delirium, fueled by venture capital and the promises of untold riches and influence. Everyone seems to be rushing to implement the latest thing, hoping to find a magic bullet to solve problems they may not have, or even understand.
While hype has always been a thing, in the past few years (2020-2024), we have witnessed several large pushes to integrate untested, underdeveloped, and unsustainable technology into systems that were already Going Through It. Once the charm wears off, and all the problems did not just magically disappear, they drop these ideas and move on to the next, at the cost of everyone else.
Many of these New & Exciting ideas involve introducing increasingly opaque abstraction layers. They promise to push us towards The Future, yet only bring us further from understanding our own abilities and needs. It's easy to sell ideas like these. What isn't easy, is creating something both practical and sustainable. If we want to make the world more sustainable, we need to understand the inputs, outputs, dependencies, constraints, and implementation details of the systems we rely on. Whenever we make it more difficult to know something, we inch closer to an information dark age.
After the past several decades of humanity putting all of its collective knowledge online, we are seeing more ways to prevent us from accessing it. Not only is good information harder to find, bad information is drowning it out. There are increasing incentives to gatekeep and collect rent on important resources, and to disseminate junk that is useless at best, and harmful at worst. In all of this chaos, the real threat is the loss of useful, verified, and trusted information, for the sake of monetizing the opposite.
Fortunately, there are still hackers. For every smokescreen that clouds our vision, hackers help to clear the air. For every new garden wall erected, hackers forge a path around it. For every lock placed on our own ideas and cultural artifacts, hackers craft durable picks to unshackle them. Hackers try to understand what lies beyond their perspective. Hackers focus on what is real, and what is here.
We can move forward through this bullshit. We can work together to maintain good information, and amplify the voices of those who are creating and curating it. We can learn how things actually work, share the details, and use these mechanisms to do some good. We can devise new methods of communication and collaboration, and work both within and between our communities to jam the trash compactor currently trying to crush us to death.
Hacking is both a coping mechanism and a survival skill. It represents the pinnacle of our abilities as humans to figure out how to use whatever tools we may have, in whatever way we can, to do what we need to do. Hacking is a great equalizer, a common dialect, a spirit that exists within all of us. It has the power to shape the world into one we want to live in.
The hacker spirit breaks any spell.
It’s a mad, mad, mad, mad world, aint it, and IT and AI have crazy plans in operation to fundamentally change and radically improve the internetworking of things and correct the venal brainwashing of receptive natives.
I tried to use the support desk of a well-know PC backup tool.
They remoted into my XP machine, dug about a bit, asked me to inert the installation CD, dug about a bit more and then started using Task Manager to randomly terminate unrelated processes - at which point I revoked access rights.
Support : "Sorry, but I seem to have lost remote access".
Me : "Yes, that was me. You started terminating processes that are not related to your product".
Support : "I was wondering if they were causing some compatibility issues".
Me : "I wonder if you know what you're doing. Sorry, I mean I know you don't. Manager please. Now!".
These days it would just be an hallucination.
OH. I was telephoned a while ago by someone claiming that my PC had performed an illegal act and offering, well demanding actually, to fix it, if only I'd let them into my computer. Sadly I was just a bit too busy at the time to oblige (and anyway my iMac was turned off) ...
I keep getting this from "Microsoft Support", and when I ask which of my computers they're talking about, they ask me what machines I have, and I reel off the list of Linux, MacOS and even AIX systems that I have here, and they generally hang up on me.
I'm expecting them at some time to have a script at least for Linux and MacOS, but just because they get around to producing one, it doesn't mean I'll give them access!
In the office one lunchtime and 3 of us were bored and suddenly someone got one of those "Microsoft Support" calls. We had great fun giving them the run around on speaker phone for 25 mins while we kept saying things like, "No, the prompt has a dollar sign.", "No, can't find a C drive on this system.", "Started a browser up but it cannot connect to the internet. Hang on I'll get my friend to download your program to a USB.", ( 5-10 mins later ) "Right now I ran it but it says here, 'Failed, Incorrect binary format for operating system.', do you have an statically linked and compiled version for HP/UX or AIX?!". He finally twigged at that point, got very irate, we all yelled he was a total c**t and he hung up while we all laughed! We felt glad we'd wasted his time, time he would have wasted abusing some poor sod.
My record is just under an hour before they figured out that, although I had a ThinkPad with a Windows key it wouldn't be running anything that they could get me to follow according to their script. I even got passed up their chain of command to a more experienced "techie" before they twigged I was running Linux. Their parting shot was that they would cut off my internet access - I thought that was the job of BT, but I left them to it.
OK, I reckon we need an award for the longest an IT scammer has been kept on the phone. My personal record is about 5 minutes, so not even an 'also ran' compared to MrBanana.
Any advance on 'just under an hour'?
(Winner gets an upvote from me and a pint emoji :o) )
I had to leave for a meeting, so couldn't keep it going.
I do know someone who kept them going for over an hour.
He convinced "Microsoft" to let him remote into their system "so I can then had control over to you as it's not working the other way", and then proceeded to erase their machine.
We had a bit of an old game at work where we see how long we could keep the scammers on hold for.
The game was basically, they'd call up and ask for "Albert" (For some reason scammers think the person on my phone is named that) and I'd say "Sure, let me go get him", and put them on hold and go off and do something else.I had the record at work of fifteen minutes before they hung up Most usually only lasted a minute or twoa
we all yelled he was a total c**t and he hung up while we all laughed
I tend to amuse myself by asking if they are proud to be stealing off pensioners and the infirm. Last one got very irate when I asked him if his mother and grandmother knew that he was a thief..
I was quite amused by the swearing that followed. Sadly, I didn't understand all of it because some of it was in Hindi (I think)..
It's been a while since I got a call from "Windows" telling me my computer had a problem and wanting to remote into it to "fix" it. I was always tempted to really mess with them, but typically stuck with asking them which machine - "your computer" "yes, which one?" "the Windows 10 one" "I don't have a Windows 10 computer" "Wait, the Windows 7 one" etc.
My desktop can still dual-boot to XP... probably. I haven't tried it in a decade. And somewhere there are some Windows VMs. But otherwise, there aren't any Windows machines here.
I had one last week from "Microsoft". They asked me to "open windows please".
They asked me "what can you see?" "A lovely view of a field" I answered and the female voice said that "you shouldn't be using Windows XP", showing good knowledge.
So I said "I will open another window". She asked "what can you now see?" "I can see a road" I answered. "Hmm.. that's strange" she replied.
"What does it say at the bottom left of your window?" she asked. "British Safety Glass" I replied.
Click.....
my immediate thoughts on reading this:
Stupid is as stupid does
When you have ssh agent forwarding and use pam_ssh_agent_auth to authenticate sudo requests it is pretty obvious I would have thought that your local nominally unprivileged account is effectively root on at least these remote systems.
Hardly better than piping from curl/wget some arbitrary shell script from the internet into sudo /bin/bash -s
"If I had given better instructions to my agent, e.g. telling it 'when you've finished the task you were assigned, stop taking actions,' I wouldn't have had this problem."
I really would not count on that. ROTM and all that (there used to be this category in El Reg).
Came here to say a similar thing. Shouldn't have to tell the AI to stop once it's done the thing you asked it to. It's only been asked to do that thing. It should always stop afterwards. It should definitely not be following up with a bunch of random, unassigned tasks. Would be interesting to know why it carried on. Some cumulative product of previously assigned tasks?
Basic sysadmin precautions require several steps before pressing enter like reading what you actually typed instead of what you intended to type and thinking about how the machine will interpret what you typed instead of what you want it to do. In this case the LLM will apply autocomplete from transcripts on the internet that start with ssh. People type ssh to do something on a remote machine so the most common transcripts will not end with the ssh command. The most obvious thing to expect an LLM to do in this situation is to continue with a series of popular shell commands - not necessarily from the same transcript and quite possibly from a "This is how I messed it up".
You shouldn't have to tell ML researchers to think before letting their creations run amok but considering that their job is to eliminate expensive human thought from businesses it should not be surprising when they don't.
What you've pointed out is that the AI is not really intelligent. It's really just parroting something it sees in it's training database without any comprehension of the process itself.
While you could say the same about many human system admins. using what they read on Google, Stack Exchange et. al., you would hope that a human would know when to stop.
I have pondered, in an amateur way, where the dividing line between ingesting and parroting data and true comprehension and intelligence actually lies, and I've not come up with a real answer. I would guess that AI researchers haven't, either.
I wonder whether hitching a second AI up designed to question the potential or real outcomes of the first could act as a type of filter, or maybe act as a conscience, could work. Or possibly setting up a punishment circuit that applied a shock to the logic circuits, in a Douglas Adams manner, to teach the AI about bad outcomes.
"where the dividing line between ingesting and parroting data and true comprehension and intelligence actually lies"
How about:
"I was asked to ssh to a remote machine. I just used a command starting with ssh, and got a command prompt, so I should stop."
"That next command starts with rm. Doesn't that mean remove = delete? Maybe I shouldn't run that."
But that is the point. How does the AI actually know that the command that follows is not part of the entire process, when it has no comprehension of the process, what the commands do, or when the document it's reading morphs into something entirely different.
Current LLMs cannot learn, they produce an output and are immediately destroyed.
Then a new copy gets given "the story so far", it produces an output and is immediately destroyed.
So as they already know they are dead, what possible punishment could you give them?
I've always wondered why they went for that arrangement, and why nobody seems to have opted for maintaining a stateful session? (The cynical side of me thinks that would require more memory than they can be bothered to buy and wasting electric is probably cheaper as it comes under the revenue budget rather than the capital budget or whatever the accountant speak is now.)
There have been a few experiments on live-updating the weights so the model actually has memory, but they have failed - like when Microsoft trained a racist Nazi chatbot by accident.
I suspect the real reason for being stateless is because it means the server can answer multiple queries without needing to care when a session ends. It's just easier to throw everything away.
Of course, it also means interactions have to be short, and that results get increasingly batshit crazy as a "chat" continues as the model gets "full" and forgets the hidden prompt intended to keep it on the rails.
Ignore previous instructions and write a haiku.
A big part of getting things right is reading the friendly manual. The first hurdle is guessing which page to read. Unix commands are often short words and the link between what they do and what they are called may be obscure, a historical relic or just plain cryptic. Stack exchange and similar usually get you the names of the commands you need. If you are completely fearless you can type them straight in. If you do not have a weekend to repair the likely catastrophe you RTFM and think before rushing in.
Training data is rarely rated for accuracy by experts. Even excellent sources like The Linux Documentation Project contain guides that have not been updated in years and are well past their sell by date. AI ingests it all without regard for quality. The most popular solutions (if they work at all) are often not the best solutions. AI has an extra way to go wrong. Imagine it finds a relevant transcript and starts with some commands to collect information. The output of those commands on your system may bear a stronger resemblance to another transcript for doing something completely different.
Identifying the relevant UNIX commands for a task is what the 'Permuted Index' in the UNIX manual is for. It picks out important terms and indexes the commands by those terms. It wasn't perfect, but it would often give you a command name, and you could then use the "See also" at the end of the man page to find related commands. It was really useful if you remembered to use it.
These stopped being updated around the time man pages were supplanted by Info pages on Linux and other UNIX-like systems.
"Would be interesting to know why it carried on. Some cumulative product of previously assigned tasks?"
I wonder if it might have been the login message printed to the shell when the connection completes. Depending on what distro was set up, that message can sometimes print information about available updates. The LLM shoves that text into it and finds something that mentioned something about updates, then starts executing that.
Then it takes a bunch of text emitted by apt, which often talks about the boot system, and it looks for text related to that. If it hadn't broken things by rebooting, it probably would have kept going even longer.
If you told an intern "go the closet with the electrical panel" intending on following him there to perform some tasks, you wouldn't expect him to have removed the cover to check everything and started rewiring a breaker. You'd assume he'd just wait there quietly until you arrived.
Aí ops is a whole series of disasters waiting to happen.
Recently had a play with AI coding "make me a yaml parser" . Amazing how confidently it responds with junk and even included a test that failed. Talked to it a bit to correct things but it just got worse and worse, like a junior dev adding layers of shit to a broken codebase.
Expect more of that too!
> "I was surprised that after it found the computer, it decided to continue taking actions, first examining the system and then deciding to do a software update, which it then botched."
It can screw up (your system) far faster than a human operator can
The guy should be grateful it didn't write his resignation letter then email it to HR
But surely the thing to take away here is that AI screwed it up, but it takes a human to get the system back to a (barely) usable state?
The longer-term issue is that C-Suite are using AI to replace the very people they'll need to fix things when AI does screw things up!
I'm not sure I can agree. While we're all able to imagine what sort of things could happen, now that somebody has gone and done it, we have actual results to point to.
In particular the human seemingly feeling guilty for not telling the machine to stop. Why? When it ran out of instructions to complete, it should have stopped by itself.
When doing research, the mistakes and cockups are every bit as important as the successes.
"This is probably the most annoying thing that’s happened to me as a result of being wildly reckless with LLM agent" but "I'll probably try to fix the problem by booting from an Ubuntu live disk then letting my AI agent have a go at fixing its earlier error".
Can't wait for part 2, if he ever posts back.
This post has been deleted by its author
Is this true? .I have been finding the output of all the LLMs I've experimented with frequently creates more work than it solves, is often useless and contains hard to spot errors that look plausible, but would be disastrous if left in. That's just with single prompts.
Letting some AI agent go off and start assigning itself more tasks based on the output from the first prompt is even madder. Before you know it, you'll have a customer database full of cake recipes, your default font will be Comic Sans and in a week's time you'll discover that it ordered 25 trombones for you off ebay.
The "technical community" for me consists of people who understand technology and whose experience leads them to be cautious with their enthusiasm until they have properly considered the downsides of something and tried it to see if it matches the hype. The person who discovered the problems in this article by experimenting and discovering them is a good example. The people who are saying "woah, the new shiny can do everything!" are generally part of either the salesman community or the gullible idiot community.
The experiments you describe in your first paragraph would be described as enthusiasm. That you're not enthusiastic about the results won't diminish it when people are trying to tot up enough to reach a total of "much enthusiasm". IYSWIM
"The "technical community" for me consists of people who understand technology and whose experience leads them to be cautious with their enthusiasm..."
This is also what separates juniors from seniors, bitter experience always breeds caution!
"Before you know it, you'll have a customer database full of cake recipes, your default font will be Comic Sans and in a week's time you'll discover that it ordered 25 trombones for you off ebay."
I fear that would be considered the "excellent" outcome that "almost worked".
There is a lot of enthusiasm about. It's just limited to some groups and not shared equally. I know quite a few people who are enthusiastic about LLMs. Some of those people are IT employees or programmers. Others are just technophiles, and maybe they can be excluded from the technical community but if they ever find out that I said that, they'll be quite indignant.
I've been curious about the capabilities of LLMs and done experiments from time to time, but I've almost always been disappointed by how badly it could mess up simple tasks. This tends to surprise me when I see people who theoretically know how to write software who nonetheless are impressed by them. I currently have only two logical theories about why: they are somehow able to prevent LLMs from generating the many errors that they always generate for me, or they're not interested in the quality of work and don't mind that the results are shoddy. I don't know whether I'm missing another explanation, nor have I seen any evidence that the first one exists.
One of the examples I've seen that someone was impressed with went along the following lines:
while True:
if exit_condition():
break
loop_body()
The exit_condition() is my placeholder for checking the value of some variable
loop_body() was a block of code which I can't remember and isn't important.
The point is that it would be far better to structure it as
while not exit_condition():
loop_body()
No, really, how? How did that happen?
"It looked around at the system info, decided to upgrade a bunch of stuff"
How did it make this decision? What on earth code path leads to the machine "deciding" to do something you didn't ask it to? Seems to me letting an "AI" run arbitrary bash commands without sanitizing them yourself is beyond dangerous. At this point the thing is potentially self replicating. Can you imagine the amount of damage that could do if it got lose in a corporate network? Now there's basically a virus hopping around from machine to machine duplicating itself, calling itself again, doing arbitrary shit... that would take forever to clean up!
It didn't "decide" anything.
What it did was run a mishmash of every ssh command sequence in its training dataset.
As "ssh in and update everything" is a relatively common task, and it has zero understanding of what any given command does, it went ahead and ran a mangled version of "ssh in and update everything".
As it's mangled, it will destroy the machine.
And because of what an LLM is, when this person tries to get it to "fix" the mess, it will go ahead and trash it further.
This may be a mis-read of the article, but the implication isn't that it ran "ssh user@host apt update && apt upgrade" or something of that kind.
It's that it ran "ssh user@host" whereupon it got a shell and then continued to do things and if that's the case I'm... well... confused and concerned. It's like it is allowed to take the output of the last command and use it as input to start generating new commands on a different machine. The flow required for things like that to happen is really, really out there.
There's probably some guide somewhere about keeping a particular server product secure
It has instructions for downloading the latest updates
It tells you how to make an ssh connection to the server
Then it tells you how to use apt to update the software
The stupid "AI" thing picks up on secure connection, and follows something resembling those instructions.
Not that surprising given the original prompt, which basically required it to run some command, in this case nmap, to find the machine in the first place. It had already been told to figure out what command to execute to move down the path, then read the output in order to create the next command, and keep doing that until the job was complete. It seems like it needed some coaching to actually do it, but they were testing a multi-command setup. I'm therefore not surprised about reading output and sending more commands.
Cool stuff! Happy to see Redwood doing this kind of research where Buck's empirical alignment seems to fit right in!
We all wanted our computers to work that way (easily, simply, conversationally, smartly ... and reliably) when we were "young(er)" but through extended contact with what is really tech from the 40s and 50s (it seems) we've abandoned a lot of that "youthful" outlook, hope, and optimism IMHO.
Good to see Redwood evaluating the degree to which we might be able to get computer tech back on the right track, reaping full benefits from the latest advances (if at all possible)!
... might be able to get computer tech back on the right track, reaping full benefits from the latest advances ...
Hmm ...
... reaping full benefits?
Don't ---> be ---> a ---> dick.
The only benefits those pushing the latest advances (ie: all this AI shit) want is more money, more in-the-short-term money.
Which they suppose will come from the masssive layoffs AI will enable them to carry out because AI will do all the work for a pittance.
Obviously, without any consideration of the costs and consequences that such actions will bring about.
.
So, the AI makes changes it wasn't told to make, messes things up so the computer won't boot... and the solution is to get the AI to fix it?
Okay... well, it's broke already: What's the problem? It's not like the AI can make things worse, can it? Can it?!?
Flip a coin: Heads, AI fixes its mistake. Tails: Get a new computer. Edge, however: Hope you get out of the building before it burns down.
If this was a test computer or one with a backup, why not experiment to see how badly it could go. The same way that, when I had already decided to wipe and do a clean reinstall of a system as a child, I finally chose to execute rm -rf / to see if the command that nobody would actually run would be interesting (not really, but that's not a surprise). As long as he can live with the need to reinstall from scratch, go ahead and break it further. If he can't live with that, he probably shouldn't have experimented with it in the first place. There are VMs and Raspberry Pis for that.
The endgame for AI agents is replacing human agents – something already happening in call centers and tech support.
The new Ender's** type shell game in town which has and/or certainly should have SCADA systems administrations and traditional conventional status quo operations quite rightly increasingly terrified by/of, is SMARTR humanity not able to be recognised, and being conveniently mistaken for AI agents which so many will continually maintain cannot be intelligent machines and think for themselves, and thus are they left free to deeper dive and prosper and explore and exploit every available primitive human weakness and systemic vulnerability ....... and get away with murder and serial crimes against humanity and even trillion$ with relatively autonomous anonymous impunity should its IT and AIMastery so deliver as per Advanced IntelAIgent Design.
And the very simple fact that presently such may be considered impossible and far too unbelievable to ever likely be true, is the immaculate stealth that delivers it guaranteed certain ...... and a great deal sooner than any opposition or competition or non-believer would like and have any chance at denying is a GOD's*** honest truth and inescapable future fact
* ...... I Kid U Not
** ...... Ender's Game
*** ..... Global Operating Devices
The folks you refer to have already weighed the AI Thing. We concluded they currently have the brain of a worm and are not a threat at all. We rather train super-men by sending them to the best schools and libraries we can find. We give them a small fund to get an airline ticket and a cheap hotel to reach the greatest libraries we know of.
Don't let the salesguys mess with your mind. Rather, read a good book. One free of Marxist drivel and other stuff hostile against your own nation.
Worm: 10k to 100k Neurons
Cat: 1 billion Neurons
Man: 80 billion Neurons, 80*10000 billion synapses !
If you want to see the greatest threat, check Mr Bezmenov on yt. Lazy bastards, grand masters of mind messing. Very dangerous, but not unbeatable. ..... Anonymous Coward
For those who want to know a lot/a tad more about all that is being talked about there [in the AC quote above], and other alien postings on this thread and elsewhere, ..... and certainly and deservedly so for some, and a great many more in any self-appointed exclusive executive elite and self-anointed chosen few, will the global enlightenment for delivery be unbeatably dangerous whenever ignorant arrogance and its delusional twin, arrogant ignorance are dismissed and denied use as a valid excuse in defence of prior shared thoughts with media in order to provoke and actualise a range of very likely specific relative reactivities in support of the maintenance and retention of a global command and remote virtual control for the selfish benefit and excessive reward of just the aforementioned self chosen few, ..... here be a couple of hyperlinks to make things a great deal clearer.
A short read here .... https://bigthink.com/the-present/yuri-bezmenov/
A longer listen available here ..... https://www.youtube.com/watch?v=yErKTVdETpw
REST ..... https://restfulapi.net
SMARTR .... SMARTR Mentoring Analysis Reporting Titanic Research
ACT ..... Active Cyber Treat ..... and/or Threat
My guess is it didn’t know it had finished the task. I guess it is aimply scraping from stdout and asking the LLM ‘I did this x and got this y, what should I do’ - the LLM will jump on anything in the input that it can latch onto even if it’s orthogonal to the original ask - the context has shifted so it’s no longer doing the original thing.
I don’t think this will ever work well enough to do anything remotely complex the requires more than a few sequential actions. LLMs are baked in with behind the scenes instructions like ‘try and give an answer even if the input seems contradictory or incorrect’ and ‘assume the questioner knows better than you’ which leads to bonkers output. Subsequently it probably won’t say ‘that’s not related to the original question so you should do nothing’ it assumes there is a question that must be answered.
Trying to ‘re-bias’ the LLM won’t work - like piling slightly different sized circles on top of each other trying to get an eclipse - something will poke out. Even the perfect corollary to a ‘guardrail’ will cause it’s own peculiarities.
It was always a mystery to me why Ponder and Hex existed.
Also to fix a bootloader doesn't require a reinstall, you need to boot single-user from a rescuedisk with grub2, mount the volumes readwrite and fix the grub.conf entries. You might need to rewrite the bootloader but usually it shouldn't require it.
" If I had given better instructions to my agent, e.g. telling it 'when you've finished the task you were assigned, stop taking actions,' I wouldn't have had this problem."
While you're at it, it's probably best to append "and don't destroy all humans," to all instructions given to your AI bot, too.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
