Schrödinger’s AI: It’s so smart it can build perfect security, but it’s too dumb to figure out how to break it.
Ron Howard narrator: Actually, they would need more.
The look on her face in the thumbnail matches the title perfectly.
AI is opening so many security HOLES. Its not solving shit. AI browsers and MCP connectors are wild west security nightmares. And that’s before you even trust any code these things write.
At this point, they’re just rage baiting and saying random shit to squeeze that bubble before it bursts.
Who is paying her?
AI might pull her head our of her ass… eventually.
At this point we need to pull their heads out of our asses
I have worked as a pentester and eventually a Red Team lead before leaving foe gamedev, and oh god this is so horrifiying to read.
The state of the industry was alredy extremely depressing, which is why I left. Even without all of this AI craze, the fact that I was able to get from a junior to Red Team Lead, in a corporation with hundreds of employees, in a span of 4 years is already fucked up, solely because Red Teaming was starting to be a buzz word, and I had passion for the field and for Shadowrun while also being good at presentations that customers liked.
When I got into the team, the “inhouse custom malware” was a web server with a script that pools it for commands to run with cmd.exe. It had a pretty involved custom obfuscation, but it took me lile two engagements and the guy responsible for it to leave before I even (during my own research) found out that WinAPI is a thing, and that you actually should run stuff from memory and why. And I was just a junior at the time, and this “revelation” got me eventually a unofficial RT Lead position, with 2 MDs per month for learning and internal development, rest had to be on engagements.
And even then, we were able to do kind of OK in engagements, because the customers didn’t know and also didn’t care. I was always able to come up with “lessons learned”, and we always found out some glaring sec policy issues, even with limited tools, but the thing is - they still did not care. We reported something, and two years ago they still had the same bruteforcable kerberos tickets. It already felt like the industry is just a scam done for appearances, and if it’s now just AIs talking to the AIs then, well, I don’t think much would change.
But it sucks. I love offensive security, it was really interresting few years of my carreer, but ot was so sad to do, if you wanted to do it well :(
Seeing all these AI ideas, i think security is about to get hugely more important in the near future.
Definitely, but the issue is that even the security companies that actually do the assesments also seem to be heavily transitioning towards AI.
To be fair, in some cases, ML is actually really good (i.e in EDRs. Bypassing a ML-trained EDR is really annoying, since you can’t easily see what was it that triggered the detection, and that’s good), and that will carry most of the prevention and compensate for the vulnerable and buggy software. A good EDR and WAF can stop a lot. That is, assuming you can afford such an EDR, AV won’t do shit - but unless we get another Wannacry, no-one cares that a few dozen of people got hacked through random game/app, “it’s probably their fault for installing random crap anyway”.
I’ve also already seen a lot of people either writing reports with, or building whole tools that run “agentic penetration tests”. So, instead of a Nessus scan, or an actual Red Teamer building a scenario themselves, you get a LLM to write and decide a random course of action, and they just trust the results.
Most of the cybersecurity SaaS corporates didn’t care about the quality of the work before, just like the companies that are actually getting the services didn’t care (but had to check a checkbox). There’s not really an incentive for them to do so, worst case you get into a finger-pointing scenario (“We did have it pentested” -> “But our contract says that we can’t 100% find everything, and this wasn’t found because XYZ… Here’s a report with our methodology that we did everything right”), or the modern equivalent of “It was the AI’s fault”, maybe get a slap on the wrist, but I think that it will not get more important, but way, way more depressing than it already was three years ago.
I’d estimate it will take around a decade of unusable software and dozens of extremely major security breaches before any of the large corporations (on any side) concedes that AI was really, really stupid idea. And at that time they’ll probably also realize that they can just get away with buggy vulnerable software and not care, since breaches will be pretty common place, and probably won’t affect larger companies with good (and expensive) frontline mitigation tools.
Execs and managers showing Dunning-Kruger in full effect.
I tried using AI in my rust project and gave up on letting it write code. It does quite alright in python, but rust is still too niche for it. Imagine trying to write zig or Haskell, it would make a terrible mess of it.
Security is an afterthought in 99.99% of code. AI barely has anything to learn from.
Mitchell Hashimoto writes a lot of Zig with AI (and this interview is almost a year old), see: https://www.youtube.com/watch?v=YQnz7L6x068&t=490s How long since you have tried tools? I think there has been some pretty astounding progress during the last couple of months. Until recently i did not use it daily, but now I just cant ignore the efficiency boost it gives me. There are definitely security concerns, and at this point you should not trust code that you do not read/understand, but tbh. i’m starting to believe that AI might (at least in the short term) free up resources to patch stuff and implement security features, that otherwise was not prioritised before due to focus on feature development. What it does to the IT sector in the long run - who knows…
If you’re using Hannah Montana Linux you can just open a terminal and type “write me ____ in the language ____” and the Hannai Montanai will produce perfectly working code every time.
Still not sure if joking.
Hannah Montana Linux is serious business. I would never joke about Hannah Montana Linux.
If this isn’t real then someone is going to make it
It does quite alright in python
That’s cause python is the most forgiving language you could write in. You could drop entire pages of garbage into a script and it would figure out a way to run properly.
Even in Python you have to keep it siloed. You have to drip feed it pieces because if you give it the whole script it’ll eat comments, straight up chop out pieces so you end up with something like
def myFunction(): # ...start of your function here...replacing actual code.
As usual, the biggest advocates for AI are the ones who understand its limitations the least.
Ha ha ha ha ha!
Oh wait, you’re serious. Let me laugh even harder.
HA HA HA HA HA!
couldn’t ai, then also, break code faster than we could fix it ?
AI WRITES broken code. Exploiting is is even easier.
How do you exploit that which is too broken to run?
Self-exploiting code.
I mean, at a high level it is very much the concept of ICE from Gibson et al back in the day.
Intrusion Countermeasures Electronics. The idea that you have code that is constantly changing and updating based upon external stimuli. A particularly talented hacker, or AI, can potentially bypass it but it is a very system/mental intensive process and the stronger the ICE, the stronger the tools need to be.
In the context of AI on both sides? Higher quality models backed by big ass expensive rigs on one side should work for anything short of a state level actor… if your models are good (big ol’ “if” that).
Which then gets into the idea of Black ICE that is actively antagonistic towards those who are detected as attempting to bypass it. In the books it would fry brains. In the modern day it isn’t overly dissimilar from how so many VPN controlled IPs are just outright blocked from services and there is always the risk of getting banned because your wifi coffee maker is part of a botnet.
But it is also not hard to imagine a world where a counter-DDOS or hack is run. Or a message is sent to the guy in the basement of the datacenter to go unplug that rack and provide the contact information of whoever was using it.
In the context of AI on both sides? Higher quality models backed by big ass expensive rigs on one side should work for anything short of a state level actor… if your models are good (big ol’ “if” that).
Turns out Harlan Ellison was a goddamn prophet when he wrote I Have No Mouth And I Must Scream.
I have no clue how you think these two are related in any way, except for the word “AI” occurring in both.
Tbf, every day that goes by is starting to feel more and more like we’re all being being tortured by a psychotic omnipotent AI… With a really boring sense of humor.
AI should start breaking code much sooner than it can start fixing it.
Maybe breaking isn’t even far, because the AI can be wrong 90% of the time and still be successful.
A few years back someone made virus that connected to an llm server and kept finding ways to infect computers in the simulated network. I think it was kind of successful. Not viable for a virus though, but an interesting idea non the less
Genius strategy:
- Replace Juniors
- Old nerds knowing stuff die out
- Now nobody knows anything about programming and security
- Everything’s now a battle between LLMs
This is a generalized problem. It’s not only programming. The world faces a critical collapse of expertise if we defer to AI.
I’ve already had to reverse engineer shitty old spaghetti code written by people who didn’t know what they were doing, so I could fix obscure bugs.
I can wait until I have to do the same thing for AI generated code.
Just apply for any top company, especially Shit Valley. They’ll 100% use GenML.
I’ve already had to reverse engineer shitty old spaghetti code written by people who didn’t know what they were doing, so I could fix obscure bugs.
I can wait until I have to do the same thing for AI generated code.
If it’s good enough for COBOL…
Fix what code? The code it broke or wrote like shit in the first place?
