Yeah it’s a bad article. Blaming AI for helping both hackers and security teams find vulnerabilities is ostrich thinking. It gets the villain right but the crime wrong.
Finding security vulnerabilities is one of the few actually legitimate benefits of AI. The fact that it’s exposing companies for having a terrible track record of building broken products is like blaming the smoke detector for going off when people keep setting the kitchen on fire.
The root cause problem is the shoddy “move fast and break things” attitude of so many companies these days. AI’s role has been to let them move even faster and break even more things.
It’s like when a kid keeps going too fast on his bike and breaks his arm, so we buy him a Ducati 999.
The recent NoisyTCP vulnerability had been latent for 11 years. Yeah there’s s lot of AI slop coding out there but not in every case. AI vulnerability scanners are going to leave no stone unturned and eventually that will be a good thing. In the meantime, we all need to brace ourselves for a bumpy ride.
Oh yeah there’s no question they’ve been finding legit bugs, including very old ones that lay undiscovered.
I think there’s a systemic issue lurking in the shadows though, and AI vibe coding is a major contributor to it: large system errors where you can’t point to any particular block of code as responsible. These issues can crop up as the result of many interacting components operating at scale. Cascading failures where many components are behaving correctly when viewed individually but the system as a whole is not.
I’ve tasked LLMs with writing quite a few small-to-medium-sized applications now, and once you get above the level of “programs that could be a single complicated line of perl script”, they absolutely need architectural guidance or they will churn out thousands of lines of crap, complete with working tests that “proves” every component is working as intended.
A few prompts of “No, no, make it like this” at the start, with occasional directed refactorings along the way does wonders for their output. But you need to know where you’re heading and how to get there, otherwise results are definitely not guaranteed.
Yeah it’s a bad article. Blaming AI for helping both hackers and security teams find vulnerabilities is ostrich thinking. It gets the villain right but the crime wrong.
Finding security vulnerabilities is one of the few actually legitimate benefits of AI. The fact that it’s exposing companies for having a terrible track record of building broken products is like blaming the smoke detector for going off when people keep setting the kitchen on fire.
The root cause problem is the shoddy “move fast and break things” attitude of so many companies these days. AI’s role has been to let them move even faster and break even more things.
It’s like when a kid keeps going too fast on his bike and breaks his arm, so we buy him a Ducati 999.
The recent NoisyTCP vulnerability had been latent for 11 years. Yeah there’s s lot of AI slop coding out there but not in every case. AI vulnerability scanners are going to leave no stone unturned and eventually that will be a good thing. In the meantime, we all need to brace ourselves for a bumpy ride.
Oh yeah there’s no question they’ve been finding legit bugs, including very old ones that lay undiscovered.
I think there’s a systemic issue lurking in the shadows though, and AI vibe coding is a major contributor to it: large system errors where you can’t point to any particular block of code as responsible. These issues can crop up as the result of many interacting components operating at scale. Cascading failures where many components are behaving correctly when viewed individually but the system as a whole is not.
I’ve tasked LLMs with writing quite a few small-to-medium-sized applications now, and once you get above the level of “programs that could be a single complicated line of perl script”, they absolutely need architectural guidance or they will churn out thousands of lines of crap, complete with working tests that “proves” every component is working as intended.
A few prompts of “No, no, make it like this” at the start, with occasional directed refactorings along the way does wonders for their output. But you need to know where you’re heading and how to get there, otherwise results are definitely not guaranteed.