So "tens of millions" could be as little as $170k USD, which is available in free credits to anyone with a company and a website from most of the large AI providers.
Yes, it is hyperbole. Every big provider that I know of - Anthropic, OpenAI, Google, AWS, Microsoft - have a human verification process where they look for some proof that you're a real and likely-to-scale company. If you can pass that bar (being VC funded is the typical signal) then you can get hundreds of thousands of dollars worth of API credits. If not then you get nothing or a very small amount of goodwill credits.
You are giving them data so they can train their models on, understand your use cases and get hooked on their services. There is no reason for them to be picky. A company that has made zero dollars was on the stock market for 2 trillion this week. Money isn't real anymore unless you don't have it.
So i could literally set up a a company with no assets, set up a website for it, and get hundreds of thousands of dollar of free credits? Where do I sign up for that?
It used to be the case that you could similarly get cloud credits from the major providers.
As an example, Microsoft used to give startups up to $100k in Azure credits with very low barriers to entry, until they changed the requirements about 18 months ago.
Now, IIRC, you can get $1k really easily, up to $5k with business verification, and up to $100k if you provide proof of being funded.
So I’m not too surprised to find generous AI token credits are available.
Funny I mentioned having concerns about this on LinkedIn earlier in the year (about AI making it way too easy to hack, it will empower skiddies at a whole other level), it is only going to get worse. The question is do we really have any good options to protect all systems? Its almost like we need harshly hardened OS’ and programming languages.
I dont think you need to be hyper wealthy to pull off hacks, just enough private AI.
Oddly enough, I think security by obscurity works well against AI attackers. At least for now.
I've seen a real attack recently and the attackers are mostly just scanning for access to well-understood systems. They won't dig too far into your bespoke stuff. But if they find a Zendesk API key, your Zendesk gets dumped instantly.
So my thinking is: the more of it is unfamiliar, the less thorough the simplistic attackers will be with it. For advanced attackers, dunno what to say.
We are going to see more systems just go offline or air-gapped
The era of the open internet is over. We’ve had our systems on public networks for decades and the next era is going to look a lot more like Russia and China’s approach to the internet, but IMO more severe
Yes but will it keep growing like the closed source ones? I wouldn’t be surprised that the gap keeps growing, considering it now becomes a national security thing.
And to run them you need computers anyway. What happens if NVDA decides to lock down their cards?
The gap between open and closed is shrinking, not growing.
Even if open weight models are banned or severely restricted in the US and allies (an idea I think would be barely-disguised regulatory capture by US tech firms), other countries will keep making them, and people in the US will use them.
It's not intended to be dismissive. It's risk prioritization. Is PQC in the same bucket as a potential asteroid disrupting life on earth, or the same bucket as a tsunami that can happen at any time?
Simply because we can envision future risk doesn't mean it is advisably actionable. For the record -- I agree that hardening for PQC is appropriate but not at all costs, if for no other reason that my understanding of current QC is it is hitting hard boundaries at the toy stage.
But yes, increase the entropy and improve the algorithms used for in transit and at rest.
This is what happened in a lot of cyberpunk stories. It's even foundational in some cyberpunk tabletop settings, like Cities Without Number. Good times.
> Its almost like we need harshly hardened OS’ and programming languages.
That is my conclusion as well. The only solution to security attacks is... better security. A lot of systems that were secure yesterday are no longer secure today due to the sheer volume of attacks and speed with which AI can help hack; it seems like today's programming languages and operating systems were not built to the security standards that we need in the AI age. I would imagine languages like Rust gain favor over C in the coming years... but who knows.
We manage about 50 WordPress servers, and security is easier than ever
With $400 in Claude subscriptions:
Moved to FreePBX on ARM. Have a ZFS storage pool for hourly backups. Have a bunch of honeypots that lead to a firewall ban. Everything is updated. Tripwire works. Have a full understanding of the endpoints used. Egress firewall. SSH isn't used (machines poll for commands). All PHP has been audited and hardened. lots of other things
What I'm trying to say is that the defenders currently have the advantage.
> What I'm trying to say is that the defenders currently have the advantage.
I can't see how this necessarily follows, or if it does temporarily, how long it could last. This entirely depends on who the attackers are (and their incentives).
Whilst your systems are more secure against conventional attacks by human developers, it doesn't account for how much more power is available to those who couldn't pull off attacks at all previously, as long as they have money for compute time. i.e. you are less vulnerable to certain attacks but massively more vulnerable to others (including just outside your realm of responsibility).
A tiny drop in the ocean of money most orgs have, especially if you compare it to how much money they would spend implementing all those security features the poster mentioned themselves.
I'm assuming failure in the entire stack. Even OpenSSH (it's happened before). One of the things we did was to segment or remove all personal information from the infrastructure - user tables are on a separate server; that server watches for more read requests than normal, and orders are written only to a separate journal. Basically... unless they get persistent, unnoticed root access, then we're ok.
They'd need a lot of luck to notice that the first packet sent to their CNC will power down the entire VM. Heck, I'm assuming the Proxmox host gets hacked - same thing. One weird packet and the power is cut. (digital loggers web power switch pro)
> A lot of systems that were secure yesterday are no longer secure today due to the sheer volume of attacks and speed with which AI can help hack
Those systems weren't secure yesterday either. AI just lowered the cost to track down the problems which have been hanging around this entire time. Ideally, that also means it's faster and cheaper than ever for companies to find and fix problems with their own software too. If this is an arms race, the attackers don't have any advantage because of AI. Their advantage right now comes from the greed, lethargy, and apathy of their targets.
> Those systems weren't secure yesterday either. AI just lowered the cost to track down the problems which have been hanging around this entire time.
I.e. they were secure.
A major problem with "cyber security" field is that, unlike in every other field dealing with security and safety, many of the practitioners in digital realm operate in abstract world of mathematics, and fail to grasp the fundamentals of reality - such as, that there is no perfect security.
In real life, security is denominated in dollars - securing costs money, especially when it degrades the value of the thing it secures (reduction in functionality or process efficiency); something is secure if it's more expensive to attack than the value the attackers can reasonably gain from it, and securing doesn't make sense past that point (or to the point the thing being secured stops providing net value).
So yes, it's very much correct to say that due to costs of attacks going down, a previously secure system is insecure now.
> So yes, it's very much correct to say that due to costs of attacks going down, a previously secure system is insecure now.
It's more correct to say that previously insecure systems are being more frequently discovered and exploited now. If I always leave the door to my house unlocked, and as a result it gets broken into tomorrow, it wouldn't be right to say that it was secured today. It was never secured and was always left unlocked and vulnerable, I've just been lucky. There is no perfect security, but security that is entirely luck based isn't security at all.
If your neighborhood was safe to the point everyone felt okay with leaving doors unlocked, you'd feel everything is secure enough. When conditions changed (e.g. deterioration of policing, economic downturn, or even population increase), and burglaries become a thing in the area, it's no longer secure.
There is no absolute security. Even your door lock is piss poor security for determined attacker[0] - it's not really securing your home, it's raising the cost (time) of the attack to the point your home is not an attractive target (relative to your neighbor, or a deli nearby, or cars, or...).
--
[0] - If someone cared enough to get in, they'd go through the window. Or drill the lock out of the door. Or tear the door out of the frame. Door locks are mostly security through inconvenience, but in most places, that is secure enough.
Nope, unless your company has some enterprise rule banning the use of AI. Defenders can (and should!) use AI to defend. What I see, especially in Europe, is that some enterprises, government agencies, and even military entities are banning AI completely, so their own people can’t use it, whether it’s local (Chinese models) or Claude. I hope we hold the people making those decisions accountable...
Sure, minimal attack surface is always a good idea.
However, assuming security never is.
All you need in order to exploit the current nginx CVE is a website serving static html pages.
> The vulnerabilities allow unauthenticated remote code execution in NGINX itself, independent of the application running behind it.
> CVE-2026-123007 was confirmed on 64-bit servers over HTTPS, even when serving static files only.
This particular vulnerability resides in the HTTP2 support, which must be enabled by the server operator. Disabling such less hardened features is maybe even more important in reducing attack surface than only reasoning about the applications being run.
HTTP2 support can be quite spotty [0], and came with a much larger attack surface than 1.1. There's... Actually quite a number of server operators who have disabled it, because it has repeatedly been the cause of security problems.
> HTTP/2.0 is not a technical masterpiece. It has layering violations, inconsistencies, needless complexity, bad compromises, misses a lot of ripe opportunities, etc. I would flunk students in my (hypothetical) protocol design class if they submitted it. - Poul-Henning Kamp [1]
[0] Like CVE-2021-31618 that let you DoS Apache servers.
Yeah...but if you both actually care about minimizing attack surface, and are actually qualified in security - I'm thinking there's about a 0.001% chance that you've enabled HTTP2.
That's what the parent and adjacent comments are talking about, yes. However minimal surface area should not be confused with minimal volume, they're related but not the same.
I have been working on web applications for 15+ years and work in IT security. I am also a long-term user of 1984. I host my domains there and promote their service.
What most likely happened here is that they didn't have their software up to date. I highly doubt there was an undisclosed 0-click RCE involved in any of this. I have recently seen quite a few services I personally used that ran outdated versions of popular software and were vulnerable to some publicly disclosed RCE. I immediately notice these things, as I am auditing these types of services all the time. There is currently at least one major company in Switzerland that takes weeks to update their front end. They are vulnerable to RCEs all the time. This is nothing special currently.
Also, the talk of massive costs makes no sense. It's something they have not seen in Iceland, maybe, but it's normal, and the costs are extremely low, not millions. Probably just someone with a single Chinese AI subscription. Seeing complex attacks carried out extremely fast or with a huge number of requests doesn't indicate high costs by any means.
I have personally also recently developed such an exploit in this challenge (and won): [https://brotli2bitcoin.com/](https://brotli2bitcoin.com/). It did take expertise to set up the agents (basically letting them do the exact same thing I would have done, using the same approach), but ultimately, it took 37 minutes and, I think, would have cost me around 11 USD in tokens if billed at the API rate. I could have sped it up to maybe 5–10 minutes if I had been very dedicated and split the work further between agents. As I have a subscription, the effective cost was closer to a few cents.
I can't say if they are willingly or unwillingly trying to deflect responsibility away from themselves, but until they prove that they were hit by an undisclosed RCE, this is just poor maintenance on their part. We need to start holding companies responsible for not updating their software on time. They shouldn't be allowed to always play victim. The journalist should ask about their CI and when they last patched their software. That's what would let us judge them.
> They shouldn't be allowed to always play victim. The journalist should ask about their CI and when they last patched their software. That's what would let us judge them.
Honestly, I don't know whether to be more concerned about the hack itself or the fact there's a company who thinks it gains them prestige to be named 1984. Is it just me or does that not feel rather odd?
donno about this particular issue, but this does seem like a convenient excuse...
as a sysadmin, I'm not sure I understand the "speed" aspect - are they merely saying "we have to respond to publication of security flaws more quickly now"?
My guess would be that “speed” could also refer to the velocity by which the attackers moved through their network
I have seen certain attacks involving days and weeks where the attackers are inside the network gathering information needed to pivot. Fanning out a million dollars of frontier agents would obviously speed that process up.
For those who haven't been tracking, AI hacking harnesses like ARTEX have started to get popular by attackers. They're noisy so detectable... But doesn't change that they make multistage attacks fast, easy, and scalable. Likewise, attackers are harvesting LLM API keys, so their attacks are free. A lot of great talks at [Un]prompted last month got into different aspects of this.
On our end, we are one of the first teams to build an AI investigation harness (louie.ai) for both vibes investigating and agentic investigation automation, and our customers include some of the top incident response teams in the world, including gov agencies who get the phone call for cases like this. They are VERY busy right now :(
He is saying it is not his company's fault, because super wealth hacking, so stop the AI, because what do, when will you people get sensible?
Maybe when he writes the book about about just how this multi-million dollar unheard of sophisticated attack was accomplished, we will get some details on how they shut down his email server.
Conspiracy theory time but I might know who this is. An exchange hosted on the platform just pulled off a massive exit scam and were suspended by 1984.
It doesn't even have a peeringdb page and appears to be singlehomed to its upstream, Nova, which is one of the more serious ISPs in Iceland with capacity in submarine cables going east/west.
And they exited with $13M in XMR, now mixed into ETH while doing so. So I'd also argue that was the reason. The founder of the exchange claiming that 1984 retained the private key and personally hacked him with the feds makes it look like there's more to the truth than the founder wants to admit publicly.
I mean, the police in Iceland pulling off something like this? You sure, buddy?
In any case, leaving private keys for anything on a server should be considered foolish practice by now. Ronins are always exploitable, it's just a question of when, not if.
And assuming any XMR mixing service to have honest and transparent values is a pretty foolish errand.
My primary DNS server has not seen a single packet from 1984's hosting infrastructure in hours, not even SOA queries. Not a big impact to me just curious what the attackers motivation might be.
TIL that 1 million Icelandic Krona is just $8K USD.
As an example, Microsoft used to give startups up to $100k in Azure credits with very low barriers to entry, until they changed the requirements about 18 months ago.
Now, IIRC, you can get $1k really easily, up to $5k with business verification, and up to $100k if you provide proof of being funded.
So I’m not too surprised to find generous AI token credits are available.
I dont think you need to be hyper wealthy to pull off hacks, just enough private AI.
I've seen a real attack recently and the attackers are mostly just scanning for access to well-understood systems. They won't dig too far into your bespoke stuff. But if they find a Zendesk API key, your Zendesk gets dumped instantly.
So my thinking is: the more of it is unfamiliar, the less thorough the simplistic attackers will be with it. For advanced attackers, dunno what to say.
The era of the open internet is over. We’ve had our systems on public networks for decades and the next era is going to look a lot more like Russia and China’s approach to the internet, but IMO more severe
Only the guys who work for some teams in the government or certain companies have access to the good stuffs.
And to run them you need computers anyway. What happens if NVDA decides to lock down their cards?
Even if open weight models are banned or severely restricted in the US and allies (an idea I think would be barely-disguised regulatory capture by US tech firms), other countries will keep making them, and people in the US will use them.
both compilers are ai technology are likely press freedom/freedom of speech.
the only answer is reducing your attack surface whether it be software or supply chain or chemistry and engineering.
it'll lean toward self limiting as we adapt, but it could get very messy. the current implementation of phones and phone systems are a huge problem.
What a shame emudhra pushed such a poor security choice. We should refer to all these *shell RCEs as being emudhraed.
Simply because we can envision future risk doesn't mean it is advisably actionable. For the record -- I agree that hardening for PQC is appropriate but not at all costs, if for no other reason that my understanding of current QC is it is hitting hard boundaries at the toy stage.
But yes, increase the entropy and improve the algorithms used for in transit and at rest.
I legit use gpg signing as a backstop against local agents going rogue and removing all my remote code.
On the other hand the article is about a business that provides online services (web hosting etc) which must be online.
That is my conclusion as well. The only solution to security attacks is... better security. A lot of systems that were secure yesterday are no longer secure today due to the sheer volume of attacks and speed with which AI can help hack; it seems like today's programming languages and operating systems were not built to the security standards that we need in the AI age. I would imagine languages like Rust gain favor over C in the coming years... but who knows.
With $400 in Claude subscriptions:
Moved to FreePBX on ARM. Have a ZFS storage pool for hourly backups. Have a bunch of honeypots that lead to a firewall ban. Everything is updated. Tripwire works. Have a full understanding of the endpoints used. Egress firewall. SSH isn't used (machines poll for commands). All PHP has been audited and hardened. lots of other things
What I'm trying to say is that the defenders currently have the advantage.
I can't see how this necessarily follows, or if it does temporarily, how long it could last. This entirely depends on who the attackers are (and their incentives).
Whilst your systems are more secure against conventional attacks by human developers, it doesn't account for how much more power is available to those who couldn't pull off attacks at all previously, as long as they have money for compute time. i.e. you are less vulnerable to certain attacks but massively more vulnerable to others (including just outside your realm of responsibility).
My opinion: LLMs are the best search engine ever. The fact that they do programming is an artifact of that.
and with that, you invite nation state/political interest oriented LLMs stream/supply chain injecting vulnerabilities and spyware into 'your' code.
1. PHP
2. WordPress
They'd need a lot of luck to notice that the first packet sent to their CNC will power down the entire VM. Heck, I'm assuming the Proxmox host gets hacked - same thing. One weird packet and the power is cut. (digital loggers web power switch pro)
Those systems weren't secure yesterday either. AI just lowered the cost to track down the problems which have been hanging around this entire time. Ideally, that also means it's faster and cheaper than ever for companies to find and fix problems with their own software too. If this is an arms race, the attackers don't have any advantage because of AI. Their advantage right now comes from the greed, lethargy, and apathy of their targets.
I.e. they were secure.
A major problem with "cyber security" field is that, unlike in every other field dealing with security and safety, many of the practitioners in digital realm operate in abstract world of mathematics, and fail to grasp the fundamentals of reality - such as, that there is no perfect security.
In real life, security is denominated in dollars - securing costs money, especially when it degrades the value of the thing it secures (reduction in functionality or process efficiency); something is secure if it's more expensive to attack than the value the attackers can reasonably gain from it, and securing doesn't make sense past that point (or to the point the thing being secured stops providing net value).
So yes, it's very much correct to say that due to costs of attacks going down, a previously secure system is insecure now.
It's more correct to say that previously insecure systems are being more frequently discovered and exploited now. If I always leave the door to my house unlocked, and as a result it gets broken into tomorrow, it wouldn't be right to say that it was secured today. It was never secured and was always left unlocked and vulnerable, I've just been lucky. There is no perfect security, but security that is entirely luck based isn't security at all.
If your neighborhood was safe to the point everyone felt okay with leaving doors unlocked, you'd feel everything is secure enough. When conditions changed (e.g. deterioration of policing, economic downturn, or even population increase), and burglaries become a thing in the area, it's no longer secure.
There is no absolute security. Even your door lock is piss poor security for determined attacker[0] - it's not really securing your home, it's raising the cost (time) of the attack to the point your home is not an attractive target (relative to your neighbor, or a deli nearby, or cars, or...).
--
[0] - If someone cared enough to get in, they'd go through the window. Or drill the lock out of the door. Or tear the door out of the frame. Door locks are mostly security through inconvenience, but in most places, that is secure enough.
You made the most simple application you could instead of the most advanced one.
Welcome back pure nginx serving static html pages as the default. Good bye 5000 dependencies.
However, assuming security never is.
All you need in order to exploit the current nginx CVE is a website serving static html pages.
> The vulnerabilities allow unauthenticated remote code execution in NGINX itself, independent of the application running behind it.
> CVE-2026-123007 was confirmed on 64-bit servers over HTTPS, even when serving static files only.
This particular vulnerability resides in the HTTP2 support, which must be enabled by the server operator. Disabling such less hardened features is maybe even more important in reducing attack surface than only reasoning about the applications being run.
https://research.jfrog.com/post/nginx2shell-all-you-need-to-...
> HTTP/2.0 is not a technical masterpiece. It has layering violations, inconsistencies, needless complexity, bad compromises, misses a lot of ripe opportunities, etc. I would flunk students in my (hypothetical) protocol design class if they submitted it. - Poul-Henning Kamp [1]
[0] Like CVE-2021-31618 that let you DoS Apache servers.
[1] https://queue.acm.org/doi/10.1145/2732266.2716278
What most likely happened here is that they didn't have their software up to date. I highly doubt there was an undisclosed 0-click RCE involved in any of this. I have recently seen quite a few services I personally used that ran outdated versions of popular software and were vulnerable to some publicly disclosed RCE. I immediately notice these things, as I am auditing these types of services all the time. There is currently at least one major company in Switzerland that takes weeks to update their front end. They are vulnerable to RCEs all the time. This is nothing special currently.
Also, the talk of massive costs makes no sense. It's something they have not seen in Iceland, maybe, but it's normal, and the costs are extremely low, not millions. Probably just someone with a single Chinese AI subscription. Seeing complex attacks carried out extremely fast or with a huge number of requests doesn't indicate high costs by any means.
I have personally also recently developed such an exploit in this challenge (and won): [https://brotli2bitcoin.com/](https://brotli2bitcoin.com/). It did take expertise to set up the agents (basically letting them do the exact same thing I would have done, using the same approach), but ultimately, it took 37 minutes and, I think, would have cost me around 11 USD in tokens if billed at the API rate. I could have sped it up to maybe 5–10 minutes if I had been very dedicated and split the work further between agents. As I have a subscription, the effective cost was closer to a few cents.
I can't say if they are willingly or unwillingly trying to deflect responsibility away from themselves, but until they prove that they were hit by an undisclosed RCE, this is just poor maintenance on their part. We need to start holding companies responsible for not updating their software on time. They shouldn't be allowed to always play victim. The journalist should ask about their CI and when they last patched their software. That's what would let us judge them.
Millions in Krona, not USD.
1USD is about 125 Icelandic Krona.
15,000 USD would be 2mil over on the other side. Which makes that a much more likely story.
Well said
as a sysadmin, I'm not sure I understand the "speed" aspect - are they merely saying "we have to respond to publication of security flaws more quickly now"?
I have seen certain attacks involving days and weeks where the attackers are inside the network gathering information needed to pivot. Fanning out a million dollars of frontier agents would obviously speed that process up.
1-2 emergency, 3am, bypass-the-safety-procedures express roll outs per year is sustainable.
1-2 a week is not.
On our end, we are one of the first teams to build an AI investigation harness (louie.ai) for both vibes investigating and agentic investigation automation, and our customers include some of the top incident response teams in the world, including gov agencies who get the phone call for cases like this. They are VERY busy right now :(
Maybe when he writes the book about about just how this multi-million dollar unheard of sophisticated attack was accomplished, we will get some details on how they shut down his email server.
Might be filters your end (or an accidental "already hit the uparrow"?)
https://bgp.he.net/AS44925
It doesn't even have a peeringdb page and appears to be singlehomed to its upstream, Nova, which is one of the more serious ISPs in Iceland with capacity in submarine cables going east/west.
I mean, the police in Iceland pulling off something like this? You sure, buddy?
In any case, leaving private keys for anything on a server should be considered foolish practice by now. Ronins are always exploitable, it's just a question of when, not if.
And assuming any XMR mixing service to have honest and transparent values is a pretty foolish errand.