If it was valuable a human would have been able to communicate it in a few lines of text. Using a LLM to blast slop devalues everything and communicates little. A drunken twitter rant has more ability to convey information than sloppy blog posts with 3 pieces of information per sentence and a load bearing attitude.
At first, there were just big logic blocks and storage that did as they were told.
Later, in hardware, there were all sorts of optimizations built in where the logic would do its own thing -- so you'd get caches and speculative execution and even on-the-fly conversion from one instruction set to another.
This abstraction was also piled onto the runtime -- you'd write machine code, then assembly, then higher level languages; eventually you'd run in interpreted or just-in-time generated bytecode (hello to Java from UCSD p-System from the 70s and Burroughs machines in the 60s). And of course... operating systems...
Now, we've got machines reading source code (and machine code, and everything in between) and they've got some ability to thing the thing to make it thing. Sometimes correctly (enough to pass existing tests) and sometimes without the ability to express to the human what the thing they thinged was.
But, we're all living in the land of mel and for some it's always been better to go fishing than to pretend to participate.
LLMs currently don’t learn, at least in the way most people mean, but I assume you mean the trainers will incorporate the knowledge in future releases.
Glad to see this is the top comment. This is written by Claude, I know because I talk to Claude for 8+ hours a day, and it says stuff like this. I stopped reading the article because of it.
Not necessarily. Only if you assume that no new insecure code is added. If ai found 1 security bugs in old code but ease of code generation with ai adds 2 new security issues (in unrelated code), then we are getting less secure.
Is it a given that it's as good at outputting secure code as it finding insecure code? That leaves a fair amount of room for people not baking in deep review steps. Also mentioned elsewhere, but if velocity is way up...that ramps up bugs anyway.
This is true for human written code too. This is why in the future all written code MUST be run through an internal security sweep before being given to your harness for writing. That's just basic due diligence at this point.
This is something agents already do as part of their normal work. Providing code that you know is vulnerable is in fact a dereliction of duty and negligence. Don't provide a POC. Simply that as it produces code a background agent just checks it without showing the work and points out potential problems.
Yes but LLMs famously produce vastly more code than humans. My hot take: LLMs will produce more insecure code than they can catch, and will also find ways to smuggle insecure code to avoid extra work.
Do we have any evidence around the rate that AI emits security bugs vs humans previously? You would think that perhaps AI would be better about this, but perhaps not.
I'm making no claims. I'm just saying that contributing before ai was difficult so people rarely did it. Now that contributing is easy, more code will be added. More code = more bugs. Always.
Sure, some projects are trying to be responsible or are trying to keep waves of slop PR's from drowning them, but that is by no means an indication of a general distaste for AI code. It's been largely embraced.
I've run Daybreak against my human written pre-AI code and found so much stuff that I don't trust human or AI written code. It all MUST be audited by a security sweep AI.
Even if Fable and Astra are superhuman at identifying and fixing bugs that does nothing when projects get bugslopped faster and cheaper by some low-end model.
I hate these meta conversations on crunchy technical topics and feel like we should be digging into the KVM security model more than this general security prognostication stuff. Having said that, I will hypocritically further this debate by pointing out that the determinant of software security moving forward is not so whether AI is drastically better than humans at avoiding common vulnerability patterns, (or instead just a little better). Rather, it's that AI will make it much easier to move subsystems or whole projects to memory-safe programming languages and towards formal verification.
...if the pace of both fixing and finding bugs got accelerated, all we're in for is machines being on average more vulnerable (because patching takes time and effort and more bugs increases that)
Don't disable nested, it's still better security than everything else, and having it on means access to real hardware virtualization. Second, the article is AI-written. It really should be rewritten by a human.
> I have yet to hear a compelling use case for nested virtualization.
AWS instances are slow to launch; they take whole seconds. Nested virtualization allows running something like firecracker/crosvm/rust-vmm atop a normal AWS instance, without needing a bare-metal one.
If you're segregating internal workloads, by all means use nested virtualization on EC2 or DO or something. Your threat model is different and you probably have a lot more room to maneuver when a KVM escape is announced, because you have some level of trust in all your workloads.
If you're running multitenant for mutually-untrusting customers, that excuse goes out the window.
One distinction I'd make as well is "are you running more than one customer's workload on the same system" (yikes, don't), vs "are you using nested virt as additional protection rather than running the customer workload directly on an instance, but one instance only gets jobs from one customer".
Another good point, but this one is maybe a little less dispositive than it looks like: it's very difficult to engineer an orchestration system that reliably segregates worker physicals from each other, so I'd rate the likelihood of a game-over situation after a guest->host escape as very high regardless of how yuo schedule workloads. It's sort of the same situation as how SSRF's are usually game-overs (you're pivoting from a hardened attack surface to a much softer one) but even harder to engineer again.
That's fair. You'd want a system designed around the notion that every customer instance must be isolated from each other without the nested virtualization.
(Aside: my current guess at the appropriate security posture towards Nitro Sandboxes, in the current era of vulnerabilities, is "decent, but probably has security holes too, and is less widely scrutinized, don't count on it either". Thoughts?)
Any recommendations for good guides on that, besides "design very careful IAM policies and security groups, harden any common services they have access to, and ideally run them under separate AWS accounts"?
Do your lambda functions always get a per-customer dedicated instance or is firecracker distributing them so that multiple customers can have VMs on the same actual box?
Sounds like it would be pretty expensive for Amazon to dedicate an instance to every customer, especially for those with small workloads.
If that's the case, why would it not hold for Amazon but would hold for everyone else?
Well this sucks. So this needs a kernel fix then? Looks like I can disable nested passthrough manually, but am I stuck waiting for downsteam fixes at this point then?
Honestly, this gives me pause on the whole "VM is sufficient for any agentic sandbox environment" theory.
Another nested KVM bug. We debated enabling nested KVM for many years, and held the line every time, because the only post-2020 KVM vulnerability we knew of required nested VMs. That decision aged well: we've gotten at least 2 more nested KVM escapes in the last year.
Certain Windows features need it unfortunately. Virtualization-Based Security (VBS), Credential Guard, Hyper-V (obviously), and I'm going to say WSL2(?)
Right, sorry, that's a fair point (and Josh Triplett made another fair point earlier). I'm sort of implicitly locked in to the "hosting company" use case here: services running mutually-untrusting isolated workloads for customers.
Mitigations are layered, first not using a shared VPS, and instead running your own bare metal server with something like proxmox to run the containerization.
Those containers could be susceptible to issues between them too, but you at least control the environment and can put more checks and balances in place.
Shared cloud and vps is going to become bumpy and more insecure the next little while. There's no way around it.
Also in this case, a "missed potential" if it eventually fails to reach anyone.
I guess its time for all of us to step aside, leave space for somebody who knowns what they are doing, and go fishing...
At first, there were just big logic blocks and storage that did as they were told.
Later, in hardware, there were all sorts of optimizations built in where the logic would do its own thing -- so you'd get caches and speculative execution and even on-the-fly conversion from one instruction set to another.
This abstraction was also piled onto the runtime -- you'd write machine code, then assembly, then higher level languages; eventually you'd run in interpreted or just-in-time generated bytecode (hello to Java from UCSD p-System from the 70s and Burroughs machines in the 60s). And of course... operating systems...
Now, we've got machines reading source code (and machine code, and everything in between) and they've got some ability to thing the thing to make it thing. Sometimes correctly (enough to pass existing tests) and sometimes without the ability to express to the human what the thing they thinged was.
But, we're all living in the land of mel and for some it's always been better to go fishing than to pretend to participate.
Perhaps they are rewriting it with less slop?
The CVE itself: https://www.cve.org/CVERecord?id=CVE-2026-99312
Sure, some projects are trying to be responsible or are trying to keep waves of slop PR's from drowning them, but that is by no means an indication of a general distaste for AI code. It's been largely embraced.
CVE still listed...
Agree that the article is downright painful to read though.
AWS instances are slow to launch; they take whole seconds. Nested virtualization allows running something like firecracker/crosvm/rust-vmm atop a normal AWS instance, without needing a bare-metal one.
If you're running multitenant for mutually-untrusting customers, that excuse goes out the window.
(Aside: my current guess at the appropriate security posture towards Nitro Sandboxes, in the current era of vulnerabilities, is "decent, but probably has security holes too, and is less widely scrutinized, don't count on it either". Thoughts?)
Any recommendations for good guides on that, besides "design very careful IAM policies and security groups, harden any common services they have access to, and ideally run them under separate AWS accounts"?
Do your lambda functions always get a per-customer dedicated instance or is firecracker distributing them so that multiple customers can have VMs on the same actual box?
Sounds like it would be pretty expensive for Amazon to dedicate an instance to every customer, especially for those with small workloads.
If that's the case, why would it not hold for Amazon but would hold for everyone else?
At some point in the past they did, for pretty much exactly this reason. I'd assume they don't anymore (I last worked there a decade ago).
That does not seem to be what HN has been thinking for months.
Just don’t.
Honestly, this gives me pause on the whole "VM is sufficient for any agentic sandbox environment" theory.
Don't allow nested KVM.
https://news.ycombinator.com/item?id=48883089
In the meantime the patching will continue until security improves!
Mitigations are layered, first not using a shared VPS, and instead running your own bare metal server with something like proxmox to run the containerization.
Those containers could be susceptible to issues between them too, but you at least control the environment and can put more checks and balances in place.
Shared cloud and vps is going to become bumpy and more insecure the next little while. There's no way around it.