No, you’re holding it wrong. Give it more context and write more skills. Use the new model and add more agents. Have the AI write all your tests for you. Have the AI review itself. Wait 6 months and it will all be perfect by then.
/s obviously.
Hell yeah, Bank CEOs and the NSA never lie.
So we just vibe coding a whole country now, huh?
It
Is
Only
Going
To
Get
Worse
We’ve been saying the same thing about climate change and look what’s happend - next to nothing.
Weeee!
We’ll run the entire planet into the ground as long as we can make a handful of rich people even more rich for some reason.
This reads like bullshit. “Pushing too many patches”? What is that supposed to mean? It’s not like they are sending each patch separately. From what we know FAA outage was because they deployed a new tool, not because of vulnerabilities in the old one. Outlook outage was supposedly caused by an existing bug, not by a new fix. That leaves us with banks and NSA and the article says “But even entities with large resources, such as the nation’s top banks and the intelligence community” which again is bullshit. Everyone knows banks don’t like spending their money on IT. There might be something to this story. It’s possible that NSA and banking system outage were caused by AI but definitely not in the way this article is describing.
It would be just like a big bank to try to turn a worldwide hacking crisis into a pitch for AI.
“Pushing too many patches”? What is that supposed to mean? It’s not like they are sending each patch separately.
I’ll have to throw in a “yeah, but.” Nobody in their right mind would bundle (say) 50 patches together into a single upgrade. The chances of regression are too great, and if it happens, it makes it harder to isolate which patch caused the problem. You can mitigate that by proper testing, but in complex systems, that takes significant time and resources, and there’s countervailing time pressure to immediately deal with zero-days.
But they were also talking about scheduled maintenance, which might suggest that they still make provision to shut the system down to apply changes, rather than having spent the money and technical effort to allow them to maintain system operations when applying changes (for example, by blue/green deployments). At very least, it’s good practice to minimize downtime even when you can’t maintain operational continuity, by applying changes in the smallest possible increments.
Minimizing downtime might matter less in banks, where back-end systems often still run in batch mode, due to the banks’ reluctance to invest properly in IT (as you mentioned) and because some of the external systems they have to interface with also run in batch (e.g., overnight clearing, which is instantaneous in most of the rest of the developed world).
Tech debt from the mainframe era is the world’s biggest submerged iceberg.
The NSA outage sounded like it was a failed comms upgrade, which is probably for different causes.
Yes, but the exact quote is “They’re trying to push through a lot of patches all at once — more than the system was designed for.”
The process can have bottlenecks and can not handle enough patches properly (as you describe, issues with testing, regressions, downtimes etc.). The system doesn’t handle patches like that. It doesn’t process each patch separately. It’s not like every dev or AI agent send a patch at once and the system fails to process them.
Bank CEOs are well known for understanding how coding works.
Its either AI driven patches, or AI driven vulnerabilities wrecking your system. Pick your poison.
This is the thing, the LLM has it’s uses and looks like it is year another powerful tool to find vulnerabilities, but vibe coding is just abusing it.
It’s actually kind of crazy if you think about it.
When they specify how the code supposed to be, they want LLM hallucinate everything that they hand waved.
But then at the same time they don’t want hallucinations because those introduce bugs.
I recently saw movement combining TDD and vibe coding, I guess that’s one way that sounds like it might work, but then you are replacing wiring code with wiring code that describes what the code should do.
I’m wondering if that actually is a net positive. I guess for someone who did TDD it is as instead wiring code + test cases they only write test cases. Somehow it reminds me of prolog.
When you TDD, you are supposed to write the minimum possible change that would result in your new test passing. A slop machine is going to try to write code that looks like the end goal of the application you’re writing. Which means it will have tons of untested unexamined side effects.
Also, of course, while you’re human-writing the code for TDD you’re supposed to be thinking about how everything currently works and what edge cases might exist, etc, that is supposed to guide you towards the next best test to write.
Well, I heard about it but didn’t try, and it sounded plausible.
What you are saying sounds like that’s where the gotcha is.
vibe coding is just abusing it.
What’s the alternative? OpenSpec or something that also won’t solve it? Honestly, this is like when everyone tried to pretend that all bad software must not have been “agile” enough. This shit is going too fast for anyone to understand and it turns out having people slowing down things with padded estimates was helping keep the systems working.
this is like when everyone tried to pretend that all bad software must not have been “agile” enough
Yeah. Nobody who writes complex, critical software uses an agile methodology. There are no databases, no operating systems, no medical devices, no industrial control systems, no large-sale telco code bases that are developed using the agile paradigm.
Agile has a sweet spot, and one attribute of that is that the blast radius is small and the cost of failure low. If you can’t tolerate “fail early, fail often,” you probably need more guardrails. Anyway, agility is a continuum, not an either/or choice.
The fixes, however, temporarily overwhelmed the systems. […] “They’re trying to push through a lot of patches all at once — more than the system was designed for.”

"They’re trying to push through a lot of patches all at once — more than the system was designed for.” That’s not how patching works…
“they’re pushing new technology on top of really old computer systems, and it just seems to overwhelm the system.” Now you have it. Old hardware and new software often don’t mix. A tale as old as time.
They’ll say anything except that they’re responsible for their own decision making.
The slop will be all-encompassing!
tech outages were caused by attempts to fix vulnerabilities identified by artificial intelligence
Umm…
“This is not an AI failure, but actually because of how excellent AI is at identifying bugs”
‘OK, and how were those patches written?’
“No more questions. I have to go now”
Did AI identify the vulnerability or did AI create the vulnerability?
Schrödinger’s code
Even if it only identified the vulnerability, you can bet that AI wrote the patch that broke everything.
If you were about to say, “it’s almost like ‘AI’ isn’t really helpful”, you’d be right.
When I read the explanation given by Microsoft, I immediately thought of AI written slopware. Locking and race condions is something AIs are particularly bad at, even worse than human programmers, who already have a bad record with this classes of bugs.
Well who could have possibly seen that coming!?
everyone
Surprised? No? Me neither.
And there it is.










