Hacker Newsnew | past | comments | ask | show | jobs | submit | shepherdjerred's commentslogin

It feels to me the only way to beat LLMs finding vulnerabilities is to have LLMs (or LLMs paired with other AI methods) constantly monitor your infrastructure, traffic, database, etc. to ensure that nothing out of the ordinary is happening

I think so too; I think apart from pushing vulnerability research forward (which is important on its own merits just for computer science reasons), the major engineering challenge of this moment is figuring out how to make platforms resilient without rewriting everything; that could be better live-patching, it could be some weird monitoring thing, it could be automatic segregation and workload isolation.

It's not "just keep up with the patches". That's going to be a losing strat.


maybe today, you cant keep up, 2 year later, AI rewrite everything safe in rust, you wont need patch everday

Hahahahaha

I'm not sure how cost effective or even viable that option would be. I think the only practical defense is some kind of baseline hardening at the OS or application level. Something like eBPF rules i.e. what Cilium Tetragon policies provide. Basically it seems like there's no way to guarantee that your application code can't be exploited in some novel way so your best option is to restrict them as much as possible.

I'm really just talking as a layman here.

My feeling is that if we use traditional security approaches that seems like a losing strategy. You cannot just keep patching. I mean obviously you should, but that is not enough.

LLMs are finding zero days in operating systems, language runtimes, everything. They will definitely be finding holes in your API. It is not a question of 'if'.


I agree that patching is just plugging holes in the dam with your fingers. I think the only viable defense is something more low level in the kernel. SELinux, eBPF, or something new and more sophisticated than those. A lightweight agent passively monitoring system calls and detecting anomalies perhaps?

And what would that solve?

> there's no way to guarantee that your application code can't be exploited in some novel way

As a thought experiment if this hypothetical application code were written in something resembling haskell I suspect it would actually prove fairly difficult if not impossible to find novel ways to exploit it. I think at that point an attackers only options are logical inconsistencies in the underlying business logic (see that one NASA rocket) or bugs in the compiler and runtime.


Or use defense in depth. There will be a vulnerability in every layer with some non-trivial probability, but rarely in multiple layers at the same time.

When Plex got hacked and my linux server had a cryptominer installed, I had no clue where to start fixing it. I loaded up Claude to analyze the system logs, and it told me what commands to run (and why) and the infection was gone in about 10 minutes. There's no way I would have known the many things I needed to do to fix and secure the system. I'm fairly new to Linux, and without the AI I would be helpless with some of the tricky things.

The right thing to do in that case is to save whatever data you need from the server and then nuke it thoroughly, which means all the disks should be wiped, don't keep the old UEFI partition around, and reinstall. It would be preferable to skip the first step and get the data from a source that wasn't compromised, but sometimes you have bring out the hazmat suits to recover stuff from compromised systems.

This is very reactive though -- imagine you were a business. You'd be screwed. Your data is already gone.

I hope most businesses aren't running plex media servers for their home media center.

You're still helpless, since you didn't do what you should have done (format).

Yeah this was my exact experience with Daybreak Blue. I hope I don’t have any vulnerabilities because the agents sure aren’t going to help me

It’s actually really easy to get access at OpenAI and Anthropic

But, at least with OpenAI, I still got flagged when doing an audit of my infra. And the only way around that is daybreak red, which is extremely locked down


On the other hand, it has never been a better time to be someone who cares about your work.

You can pore over your app, find 10 problems and have AI fix them. You can begin using the majority of your time on QA, product fit, and architecture instead of implementation and debugging.

In the past it’s been said one person alone cannot accomplish much. That is not really true today if the individual is competent and motivated.


> On the other hand, it has never been a better time to be someone who cares about your work.

No, it has never been a worse time. If you care about your work, having the slop machine shoved down your throat by managers who (wrongly) believe it will make your work so much better is pure misery. Source: I care deeply about my work, and I'm sick to death of having to fight my manager about LLMs. I just wish he would judge my work on its merits (which he admits are good), not how I got there.


On the contrary, it's the best time to not give a shit as most if not all of what we do can be offloaded to frontier models that our companies pay absurd sums for

I'm sure that includes reviewing everything

I'm of the opinion that it's a complete waste of time to review every line of code and you should instead focus on the systems/architecture.

If I'm leading a large project would I review every line a junior engineer write? No. It's not sustainable.


I don't think you fail to understand that a junior engineer is a human.

And that it's never been a large project leader's responsibility to review any code.

It has always been strongly advised to have two pair of human eyes on every line of code, though (either through reviews or pair programming).

But now it's suddenly not sustainable to have even one pair of eyes, right?

Oh, true, we can't possibly hamper the output of AI crap.


As others said, homelabs are great for learning

That aside, I love K8s for my homelab. It is a bit of work, but well worth it IMO. It makes it easy to setup things like observability, networking, storage/backups


The harness is incredibly important when you're working in parallel with multiple agents. In this regard Claude Code is the best I've found.

Yikes

Do American models allow me to ask questions about piracy

I’ve been refused. I pasted a screenshot of a torrent list for a tv show. “Which one should I get, they all look the same”. “I won’t help you with copyright infringement”.

But the same models will tell me about the -arr stack and offer to set it up for me. You just have to frame the prompt the right way, at least for me. It can’t think it’s assisting with immediate piracy. Future theoretical piracy or for educational purposes, you’ll have better luck.


Yeah they do it’s genuinely strange

Yes, actually.

As of today opus 5.5 won't for me

That's weird because I used opus 5.5 to set up a full blown torrent->seedbox->plex streaming and storage solution for me.

Do Chinese models allow this? Tit for tat is a dumb dynamic, censoring legitimate searches is dumb no matter which “side” is doing it

That's exactly why I deleted my team's metrics, dashboards, and alerts

If your users are less technical then I think it is irresponsible to allow them to make choices that would compromise security and the data of others.

If there is a vulnerability the customer should have two options: patch or take down my app. Customers should _not_ be able to decide to defer/skip security if they are not technical enough to understand the impact.

It is great though that these scans are free for your users.


An interesting detail missing here is that there are two types of scans: “quick scans” and “deep scans.” Quick scans run automatically whenever a user publishes a project, while deep scans must be triggered manually -- except for enterprise customers.

https://docs.lovable.dev/features/security#built-in-security...


We are moving this feature to all paid tiers soon. I agree that scanning on schedule needs to be available to everyone.

> If your users are less technical then I think it is irresponsible to allow them to make choices that would compromise security and the data of others.

Counterpoint: HN about the iPhone. FWIW, I agree with the statement above, but also suggest a developer account opening up developer flags is a reasonable gate.

(Not sure how else to avoid all the "developers" clamoring for non-public betas of iOS bricking their iPhones, as the same audience that would flip a bunch of dev toggles if available to everyone like developer mode in Safari then scream how Apple should have done more to protect them when their bank account is emptied.)

The Lovable class of tools is fascinating as the value prop is explicitly you do not need to be a developer, you do not need to know what you are doing. I think that makes the responsibility clear in a way the iPhone does not.


But not needing to know what you're doing as far as programming goes != not needing to know what you're doing as far as compliance/legal responsibility goes. Someone using a tool like this still needs to run their operation in a sane manner. Like, how much responsibility does Loveable shoulder here? We don't blame a car manufacturer for most car crashes. They, of course, try to design their cars to be as safe as possible, but the fact that they endeavor to make a safe car doesn't make it explicitly their fault if I'm driving blindfolded and careen off a cliff.

This is something that makes me roll my eyes... seeing people online talking about how everyone can just make their own version of w/e software now. As though running a software business is simply a matter making the computer screen light up in a particular way.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: