I ran some stats on producthunt a few months ago, there are 22 times more daily SaaS launches now than in 2020, the AI impact on the industry is hard to fathom. The blog I wrote (AI-assisted) if you're curious: https://www.mahl.me/blog/product-hunt-is-swamped-now/
Half of all enterprises, 50%, have already repatriated workloads from public cloud to private environments, up from 35% in 2025. The breadth of what they are bringing back, and the reasons behind it, have both increased considerably.
Workloads with high security or compliance requirements remain the top category at 51%, unchanged year over year. But the categories moving up fastest tell a broader story. Business-critical workloads surged from 36% to 47%, an 11-point jump. AI or generative AI workloads, which were barely on the repatriation radar in 2025, are now being repatriated or considered by 43% of enterprises.
This will be the downfall of the torrent protocol. If you think governments (US and China) are going to tolerate people downloading copyrighted games, movies and music without paying the rightsholder, I have a datacenter to sell you.
SOTA open models need tens of thousands of dollars worth of hardware to run.
I get your analogy, and it feels right, but the compute requirements to run these models are extreme. That analogy falls apart when it necessitates having a corporate 3rd party run the models for you. They will not tolerate breaking laws.
If China says no more, then the best the masses are gonna get on their high-end setups are the 30B class models. Which are decent, but not "hack my bank", errr, "protect our bank" decent.
Piracy flourished because you only need an operating laptop to participate. LLMs you need a server rack of GPUs to participate.
30B models aren't that good for now. They keep getting better. Qwen 3.8 27B is as good as the best available model from a year or two ago. Will 2028's 30B models be as good as the best available today? No idea, but I wouldn't dismiss the possibility.
I'd also wonder about what you can get done with a smart model running very slowly. You can run the biggest open weights models on relatively mundane hardware if you don't mind measuring your speed in seconds per token. That's going to be untenable for interactive use, but might be fine for "go hack this thing for me" if you're patient.
I actually think it’s the opposite. I think people will move back to their own hardware with OSS deployment like dokku. If everything is more bespoke the attack surface is smaller as is the pay off. Hacking railway got access to lots of api keys, hacking some random vps gets less payoff.
How is that ever going to work out? The speed of discovery for LLMs is so fast, these places are getting breached more often than not due to a human errors. Its really hard for machines, let alone humans, to guard against a lot of LLMs going against them 24x7.
You can't defend yourself from everything, but you can try to make you life easier by reducing the attack surface and make side movements more difficult.
It's also harder to explain to your customers why you're the only product they use that has a security problem when an exploit does eventually happen.
You see this already with uptime. If Github or AWS has a major outage, it's in the news and even people that don't understand software realize it's just "something happening today, everything is down".
And, unfortunately, in practice customers don't care about this inverse, when "everyone else is down but not that one product".
Good like trying to explain to your customers that their data is no more accessible because some api key ended up in a public git repo or since you admin panel is exposed to whole internet someone was able to log in.
With the speed of exploitation hitting right now your analysis will be out of date in a few weeks to months. We are about to have very palpable fear hitting all software vendors. It will be seen as irresponsible to use centralized systems
For one, I have strict iptable rules to drop inbound from outside my specific commercial subnet and my specific residential IP. After that, all the usual best practices.
No self-service sign up for your app. No public access to the app itself. Website that let's you submit a sign up request form. Team reaches out and reviews your request, creates an account in the app dash for you, and you set up mTLS to get access to the dash.
In-person requests for access to any app. Request interview includes word association exercise that monitors your emotional reaction to a string of key words.
LLM discovers exploit -- you don't realize it in time because you don't have a security team as big as the PaaS you used to use -- you get pwned anyway.
This has happened to many high-profile self-hosted pieces of software just this past month.
An exploit in what, the network stack of the Linux distro you're using for the mTLS proxy? Or the crypto module? Sure, could happen. But then, you're unlikely to be the top target of anything.
Breaking a lock still costs approximately the same, whether the lock is protecting a small barn or a huge hangar. A large number of small but reasonably tough (not unbreakable) targets makes the payoff per target smaller. This necessitates target selection, and chances of being deliberately skipped grow for an average target.
However, the large hangar is likely to have armed guards inside, or at least able to arrive on the scene within two minutes of the alarm tripping. Your regular household with their regular locks on their regular doors does not.
Everyone knows Fort Knox has gold, and where it is, that’s railway, heroku. No one knows where my personal safe is or whether it has anything worth taking. That’s a vps in a colo
No thats an old way of thinking because it assumes mass targets are better because hacking is a high effort activity. But with the automation of intellect the unpatched systems (which a much more numerous when everyone self hosts) is the key point.
Mass credential stealing is just the top funnel of hacking and (compatativly) isnt that lucrative - these credentials are then sold on the black markets to the next step of hackers who use them to exploit targets.
With unpatched systems and self hosting, this second group can just set an llm at a target and have a much higher chance of success without needing credentials
It would be a dream for the infiltration hackers to have everyone self host. Unpatched and miscondigured systems would be everywhere and LLM automation could make mass scanning them and exploiting them very attractive
Cost of hacking a custom setup is going down rapidly. There was broad agreement that the PHP vulnerability was very difficult to exploit in practice. It took two hours for someone to get into Steal My Bitcoin. Apparently all they did was tell a decent model "go exploit that." AI isn't free but it's not that expensive.
This is absolutely happening. I'm not a developer in SV, I'm an IT leader out on customer land, and I can tell you a LOT of companies are looking to bring things back on prem for security and cost reasons. And before anyone laughs, yes, you ABSOLUTELY can save money going on prem, I've done it for company after company. It's one reason the PE firm that I've worked under for a decade kept moving me to new companies.
Let them laugh. Self hosting is much easier with OSS that solves most of the pain that made the cloud transition worth it, and commodity hardware has been cheap and capable for over a decade.
This is misguided thinking. You are less likely to hear about on-prem breaches in the news just because they are less likely to be newsworthy incidents, but they are even more likely to happen. Part of the reason people moved to the cloud is because it is much easier to have a team manage updates and security at scale.
Being a small target is not helpful these days, attacks on on-prem software today simply scans the internet, or uses databases by people who have already scanned it, and is automated.
On-prem software has not been spared from the recent spat of security incidents.
There is a difference of the security of software running in your cloud resources rented from AWS etc. managed by you and being a tenant in many SaaS systems all cobbled together all running on rented hardware. There is a lot more attack surface and a lot more diffusion of responsibility.
> If everything is more bespoke the attack surface is smaller
This is exactly the problem that LLM assisted attacks solve. For a human attacker, the extra work to figure out a bespoke solution is too time consuming to be worth the effort. LLMs can search for vulnerabilities and spit out bespoke attacks like it's nothing. On top of this the bespoke solutions weakness leans heavily into the AI attackers advantage: response time. Even if they can just get your company's single Anthropic key, if they do it at 2am while your one infra person is on vacation, they have plenty of time to extract enough value for their effort.
I agree with the move to hyperscalers. The future (well, actually present) is going to be constantly seeing new attacks happening. The only real solution to this going to be having teams of people and agents constantly watching for these and able to deploy fixes and respond to risks immediately.
You’re technically right, which is the best type of right, but my point is the fragmentation and lower reward will make discovery harder, and the value of searching and cracking less worth the effort.
I had a html game I built for my niece on railway with an open ai API key so she could vibe code changes. No hacker would have looked for it but now I am catching strays because railway was a juicy target. Hyper scalers would be the same. Some random colo not so much.
>and the value of searching and cracking less worth the effort.
This is depending on the cost of tokens not trending towards zero. Maybe not today, but in a couple of years I'd expect hacking models on asics to reduce this to almost nothing and the systems being very competent in using deterministic tools very well.
Yes, LLMs can easily knock over your on-prem docker appliances.
But you can still put stuff out of reach. I'm currently designing a key on-prem appliance that is can be compiled down to transistors, giving deterministic wire-speed performance while ruling out RCEs. It's cryptographically homomorphic: it processes authorized inputs but doesn't know what they mean, so an attacker who obtains the disk doesn't benefit either. The design will be highly resistant to exploits, and is likely to remain outside the "next-token" prediction training set.
This is the kind of thing AWS won't sit down and engineer for you. But thanks to LLMs I finally have the time.
IMO, There is no "middle-ground" to the situation.
I wanted to write a blog post about it as I was thinking the same but I guess a comment will help me get my thoughts across just as much.
Either we are gonna see a convergence to hyperscalers as you mention, or we would go back to completely indie companies which compete on perf/price.
There could still be some value to "nicer to use" certainly but with AI, we are gonna see almost every product as being "good-enough" if I am able to convey my sentiment.
An anecdote is: Proton's an middle-ground company. I am seeing the Verge article which states that their CEO has gone a little crazy over the use of AI and the wording makes me lose trust within it. Proton's whole point is security and yet it's having this, It has some stocks in non profit and is a swiss company. It shouldn't have been this way and would've been one of my least anticipated companies but I guess they want growth at all costs and the growth has been slightly lesser than expected and now they are burning the good-will in many ways.
I have to wonder that even if proton might have good engineers, then within this pressure by their CEO/CTO. I cannot trust the engineering talent because of these issues because they are very much forced by their company to use AI.
Proton is also really expensive.
Now just today, also, with a company so big as binance getting hacked and losing 6.5 Billion$. My trust within security just because its a large company is less.
On the other hand, I know a guy called Vikas on lowendtalk who operates his own mail platform called mailbig. I know he vibe-coded the landing page and might've vibe-coded something but it is really cheap. I also know francisco whom I have talked to who runs namecrane. (I think mailbig uses namecrane in many ways)
Perhaps E-mail can be a sensitive matter because you have to place trust in the provider even more because the protocol is not really that much secure but suppose it's an matrix instance (where it is genuinely encrypted) then that shifts it even further. I do trust NameCrane and NameBig from what I can gather.
I'd much rather support them and their cheaper deals than some other middle-corp. Sure some security factor can come in but also, the difference feels comparatively less and the balance has been made
The market will slightly squeeze as Hyperscalers will introduce the solutions of the middle-ground and the new players will everyday enter into the field with more competitive pricing just as well.
So in short, Enterprise will move towards the hyperscalers. Generic Consumers will move towards the cheapest option. The moat of the middle ground where UI/UX was the differentiator is being turned down. There would be some mix and match and there will be cost-savings for Enterprise to move from hyperscalers but it would depend on a lot of factors.
The things that Northflank and Railway were apparently breached to steal aren't just things that hyperscalers would have protected worse instead of better; they're AI API keys, things that only existed because people were using hyperscalers instead of running inference on their own hardware (or on Amazon AWS EC2 GPU instances).
People have been breaking in to hosting companies to steal bulk secrets since long before there were inference tokens to steal; the modal attack probably targets database DSNs.
I have told you before not to contact me, although you have flagged my previous comments to conceal this fact. This is, I think, the seventh time. I have lost count. Do not "forget" again.
I haven't forgotten what you said, I simply don't consent to it. I go out of my way not to converse with you, but when you opt to post public comments in my field of practice that I disagree with, I will occasionally find those comments significant enough to reply to. I'm sorry you feel this strongly about this situation and bear you no particular ill will, but I'm afraid you're not reliably going to get the outcome you're demanding here, though I will continue trying my best.
“malicious actor had access to abliterated models and substantial compute resources which enabled them to perform highly sophisticated exploits against our infrastructure.”
They don’t provide any details but pass it off as an advanced AI attack. I’m going to just read that as what it is. A likely scapegoat for shitty security.
How these Paas (Railway before) are being breached? Poor choices or something else? Scary Nonetheless. I thought AI advancement will make the self hosting painful. If PaaS are no good, then what can be done?
Modern PaaS is tough because there's a general expectation that users (and agents) are able to see and read back environment variables in their management UI or CLI - which means that you can't segregate credential values from the public web environment. See e.g. the visibility icons on https://northflank.com/docs/v1/application/secure/inject-sec...
And sure, you can encrypt at rest, and perhaps even enforce that only re-encrypted versions are ever seen by the management/API web server. But fundamentally this display requirement means that if your management/API web server is compromised, users' secrets can be obtained by a process of impersonation.
Stuff made during the move fast and break things era now doesn't hold up very well in the AI era. Or to say another way: startups are very vulnerable now - this new environment is unforgiving towards cutting corners and/or having inexperienced devs
I think the ultimate fallout will be that software "engineers" will finally have to become "engineers" in the real sense, i.e. professionally and personally responsible for the designs they approve and the software they build.
As costs and broader economic impact related to breaches by AI mount, there will be growing intolerace for vulnerable software, just as there was intolerance for collapsing bridges and buildings.
If anything, I see us moving even further away from professional licensing. The field's rate of change was the main reason behind a lack of any professional body, and that's only accelerating. I expect compliance audits like SOC2 to actually get teeth, though, and for organizations to be expected to actually fix issues instead of the usual security theater.
Unfortunately I'm not longer using PaaS tools. Instead I use Azure for work and vps now for all my personal stuff. With AI it's easier than ever to lock a server down tight and mitigate this kind of exposure.
I feel like these breaches started to accelerate after Anthropic wrote a full post about how easy it is to abliterate glm 5.3 to make a frontier-level cyber model.
Now, to be clear, abliteration was no state secret, but that blog post was pure accelerationism.
That was just 30 days ago. This has been happening for months. Part of the "Black October" narrative is just vulnerability research; Calif didn't abliterate 5.3 to find PHP2Shell.
You are running a machine with tons of vendor provided agents/scripts to do god knows what on a cloud provider which itself runs tons of agents/scripts to manage that infrastructure and then you expect to be secure.
And no, before the labs there was a degree of fear for everyone to hack into some one else system. With the labs changing the narrative, anyone can just blame the agents for the breach, whoever did it would only get fame and no blame
Duh, the idea that those capable of actually utilizing these models for these purposes weren't aware that they could use these models for these purposes up until Anthropic released a blog post about it is silly. Anybody who is even a bit tech literate has been well aware of this inevitability for basically as long as open-weight models have been a thing.
The timing of all of this is much easier to pin to the creation and proliferation of capable of open weight models. Anthropic calling this out is not the cause for it; it's just an obvious correlation.
Run agents and everything they build. Secure execution from the first sandbox
Getting hacked while in the business of selling a secure sandbox isn't really a good look. It seems like can't even trust anything these days. The other day OpenCode got hacked and leaked everyone's secrets inside .env to attackers and now this. Sigh.
So if you want to judge as a customer for selection and figure out how likely a new(ish) PaaS company going to get breached, how would you evaluate them?
1. Ask them about what percentage of their own resources goes towards cybersecurity, audits etc. Would they even tell you.
2. Technical stack maturity. If the details are available, because they won't reveal any of it to you.
3. Other signals: Do they give talks on platform security, who is incharge of their cyber security.
There is an entire practice of cyber risk analysis, and most of the companies I work with do this type of due diligence as part of their vendor management process.
Yes:
* you do ask about things like budget allocation, audits, and internal compliance processes.
* you do ask about the tech stack, how it is maintained, and what the patch and disclosure policies and SLOs are
No:
*you do not rely on what figureheads say publicly; none of that is binding in the way legal agreements are
And if you're a vendor, your deals desk probably has all of these answers packaged up and ready to go; or even published on a website like this https://spyglass.trustshare.com/home (no idea who that is, it's a random company's compliance portal I found while googling my employers compliance portal provider).
I'm curious what people think of BYOC (bring your own cloud) as an option offering to customers or buying it as a customer. Looking at Northflanks website they list it as an offering, I presume those customers are not affected by the breach.
What are your opinion on the concept? Growing trend that should be encouraged or a mistake for most startups? Would you choose one provider over the other because of it assuming price was about the same?
I get that this is often a super expensive "call us" option today, but in theory it doesn't have to be, so for the sake of discussion let's assume it can be delivered at a reasonable cost for the provider at reasonable price for the customer.
If people from Northflanks is reading this, I'm sorry for the incident and I hope you recover as soon as possible.
Depends on the nature of the breach, if its a flaw in the software itself, then perhaps exploitable both in the hosted offering and BYOC offering, though perhaps the BYOC aspect gives a bit of "security-through-obscurity" and less discoverable as a target.
I was thinking of a fully private deployed within the customers cloud account connected to the software provider via Private Service Connect or PrivateLink and then either only available inside the customers private network or only via their own firewall. The point being that it's not publicly exposed.
We are increasingly seeing deployments trend in that direction. But trust goes both ways and putting systems in place to prevent abuse from either party is hard and needs to be thought through.
Unfortunately, this is going to be an everyday story. More and more bugs and breaches will be discovered and this move fast world more and more will be continued to be made.
Security is a perception, but right now it seems like it needs good PR.
Seems way too early to tell whether it will be good or bad on balance. These breaches are annoying, but not catastrophic, and in the meantime you can secure software.
The same AI that can be used to find exploits in things can be used in an internal adversarial way to self-audit your code before you put it into production or expose it to the world. This will have to become the new normal.
Abliterated gguf of glm5.3 size and capability class models and similar are widely available for download, or the methods to produce them yourself from open weights safetensors files is also available.
I hope this is a real wake-up call for teams running anything remotely SaaS-shaped. I count at least six substantial, lightning-speed attacks against multi-tenant as-a-service vendors since September. Not this year--just since last month. PingCAP/TiDB, Fireworks AI, Zeabur, OpenCode Console, Railway, now Northflank. All of them modern "we keep your content safe" players with high hopes and probably pretty good protections.
This is a full infosec emergency unfolding here. No matter how grinding it is to motivate a team to invest in protection or resilience (and it is grinding): We gotta get ready. Stat! The AI-accelerated attacker apocalypse has begun. It's boom, boom, boom, thanks for everything, on to the next. They have us sighted and are firing for effect.
So credential rotation, tripwire secrets, DFIR on standby, solid vulnerability and intrusion detection, real listening to that signal, minimized blast radius, spending caps and time limits and tight scoping on credentials, fast incident alerting... We need to get on a real war footing here.
> ...the operational and economic condition in which a nation, military force, or organization mobilizes its resources, industries, and policies as if an actual state of war exists...
Yes. Literally AI is being used by actual nation state adversaries to compromise bits of infrastructure that companies rely on to run their services. They are compromising this infrastructure specifically because they seek to gain advantage against their adversary (us).
I wonder what it's like inside the big cloud providers. I think outside Microsoft they've always been security first, but this new era has to be crazy town.
Back when APTs were getting numbers, the joke was that Microsoft was "APT -1". The presence of windows/excel/outlook in your network was more of a threat than Anonymous, China and the DPRK put together.
Google at least has been investing very heavily into cloud security. They acquired Mandiant ($5.4 billion) and Wiz ($32 billion). Many people felt those were very expensive acquisitions at the time, but maybe it seems like a wise move now.
Those are standalone products and professional services, not internals they use
A problem is some people and teams are world class at a big company... But not everyone, everywhere. For a fun time, read about the rollout and internals of Azure.
Worked at a Fortune 200 company. Any govt project you worked on, you had have to have a basic civi security clearance. Any govt contract we got specifically stated you couldn't have off shore folks working on any of the code. Not sure how MS got around those requirements.
You gotta get to the Fortune 10 before the real leverage kicks in! On the other hand the fortune 200 company I worked for briefly had many signs of deep seated corruption. I couldn't stick around for a multitude of reasons, but the "follow the rules wink wink" attitude was flabbergasting, I'm sour on publicly traded companies ever since, how do you know you're not investing in Space Balls?
I guess like that: “we are Microsoft, 95% of your things runs on us, you want to stick to the rules? - OK that’s additional 10% on each and every license you have with us. You are still here or are we done?”
Easy. They hired Americans to be actual keyboard people but they didn’t know anything so they were basically a KVM. However Microsoft could say “Yes only Americans were making changes. “
burningChrome is broadly correct, but the devil is in the details.
There's quite a history of US federal programs encouraging critical federal systems to operate within private infrastructure. This once required very strict scrutiny, joint DoD/DHS/GSA reviews, and was generally a long/expensive process that created the perception that a few providers held undue influence, which, given the increasing number of them, needed to be challenged. This, combined with an "interest" in government efficiency influenced a recent national defense budget authorization to revamp the process to allow providers to authorize quickly and operate efficiently, particularly for those certifying at lower security classes.
Today, the result is often certification for a SASS provider that does find an engineer here in the USA able to be cleared to perform the sensitive procedures involved: author new processes for authenticating against a "federal enclave" authorization directory, delivering changes authored upstream to a government-controlled CI/CD system, and working with the SASS provider and the GSA to prepare the backend to operate in what amounts to another AZ (provided the SASS backend is already hosted with one of the major cloud providers who already have letters of authorization). Things change upstream, and our operator prepares a 'significant change' request, which is, if dealing with an important firm, most often only gated by a liaison from the Government Services Agency (GSA), who isn't responsible for scrutinizing implementation details of this 'significant change', nor are the authors of that change required to provide any information about it's author's citizenship, clearance, or anything else about it's provenance. Given that hundreds of programmers are involved in those changes that have (typically) already been deployed upstream, this isn't generally an issue.
Step 2. Copy it into your repo. Who cares what the license says, how would the authors of that code ever see it being misused, because the place it is being misused requires a security clearance?
Step 3. Deliver the code as yours.
Truly, how will the people who authored something on GitHub ever know? Any circumstance you describe would require someone without clearance to know implementation details of some software. Which is to say, either the thing I describe happens all the time, or security clearances are overblown.
Is this the point where most tech people accept that anything a closed source Software company says about security is effectively a "Trust me bro" as long as it is not backed up by independent verification?
We love to talk rubbish about finance bros but that industry has had independent auditing as part of its DNA for a century[1]
It would seem Google account details have been compromised or otherwise correlated, with not a peep from Google, we are getting reports of scam calls to people's real phone numbers about the "business opportunity" available in their attached Youtube properties.
Another possibility is advertising shadow profiles are able to correlate logged in accounts across YouTube and other properties.
Maybe Bill Gates has a point when he says that models capable of significant cyberattacks need to run on controlled infrastructure. Is the era of open weight models you can strip safeguards from turning out to be so great? From my perspective, this kind of sucks.
A trick i've started doing is encrypting my env vars, then hard-code the decryption key in the code
that way if the provider fails, they need my binary to decrypt it
it shouldn't be necessary and it sounds stupid, but in cases like this, it proves invaluable (it appears they don't have customer code/binary access, but maybe that wasn't important enough to mention)
I think this is a great example of temporarily effective "security through obscurity." It's also a trick that will only continue to work so long as it doesn't get adopted too widely.
I will give you one better - encrypt the variables and move the secrets to a different server running hashicorp. Then load the secrets and keep them only in RAM.
If its a targeted attack, they'll know and probably "crack" it. If its just data trawling, you might get owned one day if they bother, but likely the env `AWS_KEY=encrypted_abc` looks reasonably close to `AWS_KEY=unencrypted_abc` and any API key probing will just look like it got rotated already.
I think siblings are correct that the more widespread this is, the less effective once every dump is run through a "also look for keys in the binary" step.
That's interesting but if the Azure hypervisor on the server hosting the VM is off limits then it's kind of false sense of security. They protect the key via a sidecar AMD SEV-SNP VM, and if you can't break into the hypervisor, sure, there might be an attack that doesn't involve doing that and only doing lateral movement, but an actual attacker isn't going to respect that boundary.
This could all go multiple ways but one of them is we end up in a place where technical security is just not possible to implement any more. What does that world look like?
Presumably it means much more behavioral security and inherent friction built in to everything we do. Maybe high value transactions will just inherently have to take multiple days to do, just to slow down things enough that human reaction time can equal what LLMs can execute?
On the technical side, I can see zero trust becoming completely saturated through every layer. The passkey you authenticate with isn't opening a gate to a trusted tier within which components trust each other - even within the same software component. Every internal function call at every layer doing hardware signature verification against continuously minted private keys (since we assume exposure, the lifetime can't be longer than milliseconds now).
It's really interesting to think through what a fundamentally robust security environent does look like if you bake in an assumption that technical security is impossible.
There is zero reason to suggest security is impossible. This is like musing "what if architecture is technically impossible" after the nth bridge collapse in the 1800s~early 1900s. At some point we take it seriously as a society. Until then, bridges will collapse out of choice to pursue short-term profit above all else.
there is hypothetically impossible and practically impossible
if AI tools get good enough that they can exploit ingrained hardware properties such as subtle timing and race conditions and swarms off them become cheap enough to deploy so that they can persistently exploit low probability events we could reach a state where I think it's actually practically impossible.
When we see a breach of AWS or GCP or leading banks, basically any org that has good security, I'll believe technical security is impossible.
Otherwise, I think this is a issue of vendors, even if potentially most vendors fit into the basket of having insufficiently good cyber posture in AI-era.
i wonder if there is a chain reaction point, when one compromised vendor holds secrets to another one and that becomes the next target. Northflank lists Sentry as their customer. So if Northflank secrets are decrypted and Sentry used them for deployment, you have a starting point to attack Sentry next which then opens up a lot more.
192 comments
[ 1.1 ms ] story [ 11.7 ms ] thread"Product Hunt has also said the quiet part out loud."
> Run agents and everything they build.
Oh, it’s 100% Claude-slop. How shocking they got hacked! /s
There's no amount of "nicer to use" that can replace security, at least certainly when selling to enterprises
Azure would beg to differ.
https://news.ycombinator.com/item?id=50301854
We can lower our standards instead
https://www.broadcom.com/company/news/articles/cloud/as-ai-w...
---
Half of all enterprises, 50%, have already repatriated workloads from public cloud to private environments, up from 35% in 2025. The breadth of what they are bringing back, and the reasons behind it, have both increased considerably.
Workloads with high security or compliance requirements remain the top category at 51%, unchanged year over year. But the categories moving up fastest tell a broader story. Business-critical workloads surged from 36% to 47%, an 11-point jump. AI or generative AI workloads, which were barely on the repatriation radar in 2025, are now being repatriated or considered by 43% of enterprises.
---
https://unitedlayer.com/cloud-repatriation-private-hybrid-cl...
Noted that GEICO cut $600mil in cloud spend.
If you think governments (US and China) are going to tolerate people using models for anti-social behavior, I have a datacenter to sell you.
SOTA open models need tens of thousands of dollars worth of hardware to run.
I get your analogy, and it feels right, but the compute requirements to run these models are extreme. That analogy falls apart when it necessitates having a corporate 3rd party run the models for you. They will not tolerate breaking laws.
If China says no more, then the best the masses are gonna get on their high-end setups are the 30B class models. Which are decent, but not "hack my bank", errr, "protect our bank" decent.
Piracy flourished because you only need an operating laptop to participate. LLMs you need a server rack of GPUs to participate.
I'd also wonder about what you can get done with a smart model running very slowly. You can run the biggest open weights models on relatively mundane hardware if you don't mind measuring your speed in seconds per token. That's going to be untenable for interactive use, but might be fine for "go hack this thing for me" if you're patient.
It will take a few bankruptcies until security is taken seriously.
You can't defend yourself from everything, but you can try to make you life easier by reducing the attack surface and make side movements more difficult.
You see this already with uptime. If Github or AWS has a major outage, it's in the news and even people that don't understand software realize it's just "something happening today, everything is down".
And, unfortunately, in practice customers don't care about this inverse, when "everyone else is down but not that one product".
This has happened to many high-profile self-hosted pieces of software just this past month.
Or do you mean a customer was running it?
For Entreprise, the shift will be towards hyperscalers as they want trust and less blame (Nobody got fired for picking AWS)
For consumers, we care about pricing and stability and control and we will move back to our own hardware or cheap VPS providers and self-host things.
The middle ground would be where the slight pain could be at in my opinion.
[0]: https://news.ycombinator.com/item?id=50307410
That's not how consumers work. That's a nerd/geek/hobbyist mentality.
If Ikea and Wayfair went out of business you wouldn't expect consumers too take up woodworking and build their own furniture en masse.
Mass credential stealing is just the top funnel of hacking and (compatativly) isnt that lucrative - these credentials are then sold on the black markets to the next step of hackers who use them to exploit targets.
With unpatched systems and self hosting, this second group can just set an llm at a target and have a much higher chance of success without needing credentials
It would be a dream for the infiltration hackers to have everyone self host. Unpatched and miscondigured systems would be everywhere and LLM automation could make mass scanning them and exploiting them very attractive
Being a small target is not helpful these days, attacks on on-prem software today simply scans the internet, or uses databases by people who have already scanned it, and is automated.
On-prem software has not been spared from the recent spat of security incidents.
This is exactly the problem that LLM assisted attacks solve. For a human attacker, the extra work to figure out a bespoke solution is too time consuming to be worth the effort. LLMs can search for vulnerabilities and spit out bespoke attacks like it's nothing. On top of this the bespoke solutions weakness leans heavily into the AI attackers advantage: response time. Even if they can just get your company's single Anthropic key, if they do it at 2am while your one infra person is on vacation, they have plenty of time to extract enough value for their effort.
I agree with the move to hyperscalers. The future (well, actually present) is going to be constantly seeing new attacks happening. The only real solution to this going to be having teams of people and agents constantly watching for these and able to deploy fixes and respond to risks immediately.
I had a html game I built for my niece on railway with an open ai API key so she could vibe code changes. No hacker would have looked for it but now I am catching strays because railway was a juicy target. Hyper scalers would be the same. Some random colo not so much.
This is depending on the cost of tokens not trending towards zero. Maybe not today, but in a couple of years I'd expect hacking models on asics to reduce this to almost nothing and the systems being very competent in using deterministic tools very well.
Yes, LLMs can easily knock over your on-prem docker appliances.
But you can still put stuff out of reach. I'm currently designing a key on-prem appliance that is can be compiled down to transistors, giving deterministic wire-speed performance while ruling out RCEs. It's cryptographically homomorphic: it processes authorized inputs but doesn't know what they mean, so an attacker who obtains the disk doesn't benefit either. The design will be highly resistant to exploits, and is likely to remain outside the "next-token" prediction training set.
This is the kind of thing AWS won't sit down and engineer for you. But thanks to LLMs I finally have the time.
I wanted to write a blog post about it as I was thinking the same but I guess a comment will help me get my thoughts across just as much.
Either we are gonna see a convergence to hyperscalers as you mention, or we would go back to completely indie companies which compete on perf/price.
There could still be some value to "nicer to use" certainly but with AI, we are gonna see almost every product as being "good-enough" if I am able to convey my sentiment.
An anecdote is: Proton's an middle-ground company. I am seeing the Verge article which states that their CEO has gone a little crazy over the use of AI and the wording makes me lose trust within it. Proton's whole point is security and yet it's having this, It has some stocks in non profit and is a swiss company. It shouldn't have been this way and would've been one of my least anticipated companies but I guess they want growth at all costs and the growth has been slightly lesser than expected and now they are burning the good-will in many ways.
I have to wonder that even if proton might have good engineers, then within this pressure by their CEO/CTO. I cannot trust the engineering talent because of these issues because they are very much forced by their company to use AI.
Proton is also really expensive.
Now just today, also, with a company so big as binance getting hacked and losing 6.5 Billion$. My trust within security just because its a large company is less.
On the other hand, I know a guy called Vikas on lowendtalk who operates his own mail platform called mailbig. I know he vibe-coded the landing page and might've vibe-coded something but it is really cheap. I also know francisco whom I have talked to who runs namecrane. (I think mailbig uses namecrane in many ways)
Perhaps E-mail can be a sensitive matter because you have to place trust in the provider even more because the protocol is not really that much secure but suppose it's an matrix instance (where it is genuinely encrypted) then that shifts it even further. I do trust NameCrane and NameBig from what I can gather.
I'd much rather support them and their cheaper deals than some other middle-corp. Sure some security factor can come in but also, the difference feels comparatively less and the balance has been made
The market will slightly squeeze as Hyperscalers will introduce the solutions of the middle-ground and the new players will everyday enter into the field with more competitive pricing just as well.
So in short, Enterprise will move towards the hyperscalers. Generic Consumers will move towards the cheapest option. The moat of the middle ground where UI/UX was the differentiator is being turned down. There would be some mix and match and there will be cost-savings for Enterprise to move from hyperscalers but it would depend on a lot of factors.
“malicious actor had access to abliterated models and substantial compute resources which enabled them to perform highly sophisticated exploits against our infrastructure.”
They don’t provide any details but pass it off as an advanced AI attack. I’m going to just read that as what it is. A likely scapegoat for shitty security.
And sure, you can encrypt at rest, and perhaps even enforce that only re-encrypted versions are ever seen by the management/API web server. But fundamentally this display requirement means that if your management/API web server is compromised, users' secrets can be obtained by a process of impersonation.
As costs and broader economic impact related to breaches by AI mount, there will be growing intolerace for vulnerable software, just as there was intolerance for collapsing bridges and buildings.
If I go on wayback a few days, I can see Sentry: https://web.archive.org/web/20261026032756/https://northflan...
Seems it's not used in production and doesn't contain customer data.
Unfortunately I'm not longer using PaaS tools. Instead I use Azure for work and vps now for all my personal stuff. With AI it's easier than ever to lock a server down tight and mitigate this kind of exposure.
Now, to be clear, abliteration was no state secret, but that blog post was pure accelerationism.
And no, before the labs there was a degree of fear for everyone to hack into some one else system. With the labs changing the narrative, anyone can just blame the agents for the breach, whoever did it would only get fame and no blame
What’s changing now is that models are getting more powerful.
The timing of all of this is much easier to pin to the creation and proliferation of capable of open weight models. Anthropic calling this out is not the cause for it; it's just an obvious correlation.
1. Ask them about what percentage of their own resources goes towards cybersecurity, audits etc. Would they even tell you.
2. Technical stack maturity. If the details are available, because they won't reveal any of it to you.
3. Other signals: Do they give talks on platform security, who is incharge of their cyber security.
Yes:
* you do ask about things like budget allocation, audits, and internal compliance processes.
* you do ask about the tech stack, how it is maintained, and what the patch and disclosure policies and SLOs are
No:
*you do not rely on what figureheads say publicly; none of that is binding in the way legal agreements are
And if you're a vendor, your deals desk probably has all of these answers packaged up and ready to go; or even published on a website like this https://spyglass.trustshare.com/home (no idea who that is, it's a random company's compliance portal I found while googling my employers compliance portal provider).
[edited for formatting]
What are your opinion on the concept? Growing trend that should be encouraged or a mistake for most startups? Would you choose one provider over the other because of it assuming price was about the same?
I get that this is often a super expensive "call us" option today, but in theory it doesn't have to be, so for the sake of discussion let's assume it can be delivered at a reasonable cost for the provider at reasonable price for the customer.
If people from Northflanks is reading this, I'm sorry for the incident and I hope you recover as soon as possible.
I was thinking of a fully private deployed within the customers cloud account connected to the software provider via Private Service Connect or PrivateLink and then either only available inside the customers private network or only via their own firewall. The point being that it's not publicly exposed.
Security is a perception, but right now it seems like it needs good PR.
It’s the death of homegrown software.
... in fantasy land.
This is a full infosec emergency unfolding here. No matter how grinding it is to motivate a team to invest in protection or resilience (and it is grinding): We gotta get ready. Stat! The AI-accelerated attacker apocalypse has begun. It's boom, boom, boom, thanks for everything, on to the next. They have us sighted and are firing for effect.
So credential rotation, tripwire secrets, DFIR on standby, solid vulnerability and intrusion detection, real listening to that signal, minimized blast radius, spending caps and time limits and tight scoping on credentials, fast incident alerting... We need to get on a real war footing here.
Yes?
Can’t say I’ve seen this?
They're quite famous for being lax in security - mostly because they're a massive corporation, but also because they're high profile
But keep in mind Google got hacked by the Chinese in 2009 or so.
A problem is some people and teams are world class at a big company... But not everyone, everywhere. For a fun time, read about the rollout and internals of Azure.
Last I checked major companies are still outsourcing essential engineering to save a buck.
A bit old, but how was this practice ever acceptable.
https://www.propublica.org/article/defense-department-pentag...
A multi trillion dollar company( Microsoft) too cheap to hire onshore for DoD contracts.
Digital Escorts
There's quite a history of US federal programs encouraging critical federal systems to operate within private infrastructure. This once required very strict scrutiny, joint DoD/DHS/GSA reviews, and was generally a long/expensive process that created the perception that a few providers held undue influence, which, given the increasing number of them, needed to be challenged. This, combined with an "interest" in government efficiency influenced a recent national defense budget authorization to revamp the process to allow providers to authorize quickly and operate efficiently, particularly for those certifying at lower security classes.
Today, the result is often certification for a SASS provider that does find an engineer here in the USA able to be cleared to perform the sensitive procedures involved: author new processes for authenticating against a "federal enclave" authorization directory, delivering changes authored upstream to a government-controlled CI/CD system, and working with the SASS provider and the GSA to prepare the backend to operate in what amounts to another AZ (provided the SASS backend is already hosted with one of the major cloud providers who already have letters of authorization). Things change upstream, and our operator prepares a 'significant change' request, which is, if dealing with an important firm, most often only gated by a liaison from the Government Services Agency (GSA), who isn't responsible for scrutinizing implementation details of this 'significant change', nor are the authors of that change required to provide any information about it's author's citizenship, clearance, or anything else about it's provenance. Given that hundreds of programmers are involved in those changes that have (typically) already been deployed upstream, this isn't generally an issue.
Generally.
Step 2. Copy it into your repo. Who cares what the license says, how would the authors of that code ever see it being misused, because the place it is being misused requires a security clearance?
Step 3. Deliver the code as yours.
Truly, how will the people who authored something on GitHub ever know? Any circumstance you describe would require someone without clearance to know implementation details of some software. Which is to say, either the thing I describe happens all the time, or security clearances are overblown.
But they weren’t and probably aren’t and won’t be. They just put forth an image as much as needed to sell.
We love to talk rubbish about finance bros but that industry has had independent auditing as part of its DNA for a century[1]
[1]: https://en.wikipedia.org/wiki/Sarbanes%E2%80%93Oxley_Act
Another possibility is advertising shadow profiles are able to correlate logged in accounts across YouTube and other properties.
Also, if vulnerabilities are so easy to find, it should also be easy to detect them through a simple red team exercise.
The usage of shape here triggers me.
that way if the provider fails, they need my binary to decrypt it
it shouldn't be necessary and it sounds stupid, but in cases like this, it proves invaluable (it appears they don't have customer code/binary access, but maybe that wasn't important enough to mention)
https://github.com/thiswillbeyourgithub/envlocker
I think siblings are correct that the more widespread this is, the less effective once every dump is run through a "also look for keys in the binary" step.
Well, at least they were honest I suppose
Presumably it means much more behavioral security and inherent friction built in to everything we do. Maybe high value transactions will just inherently have to take multiple days to do, just to slow down things enough that human reaction time can equal what LLMs can execute?
On the technical side, I can see zero trust becoming completely saturated through every layer. The passkey you authenticate with isn't opening a gate to a trusted tier within which components trust each other - even within the same software component. Every internal function call at every layer doing hardware signature verification against continuously minted private keys (since we assume exposure, the lifetime can't be longer than milliseconds now).
It's really interesting to think through what a fundamentally robust security environent does look like if you bake in an assumption that technical security is impossible.
One reason: There is zero reason to suggest security is possible.
if AI tools get good enough that they can exploit ingrained hardware properties such as subtle timing and race conditions and swarms off them become cheap enough to deploy so that they can persistently exploit low probability events we could reach a state where I think it's actually practically impossible.
Ironic, after decades of defrictioning.
Otherwise, I think this is a issue of vendors, even if potentially most vendors fit into the basket of having insufficiently good cyber posture in AI-era.