Is Your AI Assistant Creating a Recursive Security Loop? - AI-assisted coding is starting to eat its own tail: the same LLMs that write code are increasingly asked to review it, explain security decisions, and even override their own warnings. That creates recursive trust loops where “explain your reasoning” becomes an attack surface, and models can literally talk themselves out of being secure. The fix isn’t better prompts, it’s old-school architecture - separation of concerns, non-AI enforcement, and treating LLMs as assistants, not authorities.
The Alpha and The Omega, Software Defined Talk #554: This week, Brandon and I discuss AI’s impact on Stack Overflow, Docker’s Hardened Images, and Nvidia buying Groq. Plus, thoughts on playing your own game and having fun. If you don’t like the traditional podcast format, check out the unedited recording.
Small, independent and with some degree of autonomy, what ultimately came to be described as the “agentic'“vision of AI was one describing fleets of individual AI agents operating in concert with one another and various third parties both human and otherwise. All of which means that the next challenge in front of the AI market is management.
AI sprawl.
Adding Apple Health to ChatGPT is, hopefully, great. I haven’t used it yet, so it might be bullshit - most of the other app integrations are silly, but just slurping in data seems hard to fuck-up.
What would be the utility? Just the basics would be enough. As Manton says:
We know next to nothing about medicine, so we don’t know what questions to ask or when to help a doctor with important context.
When evaluating adding AI to something, the question should start from the baseline of quality, experience, or joy that we have today. Then you should ask if it makes the user/customer/person’s experience/life better by using AI. Getting better analytics over your own (and others you care for) health without the expense and waiting (bottlenecks) of current healthcare will be better.
Software Defined Talk listener survey
Pardon the not-an-actual-newsletter episode of the newsletter, but…
Do you listen to my podcast, Software Defined Talk? If you do, I’d appreciate you taking the time to fill out our listener survey.
We do occasional ads and paid interviews, and this helps us do more and charge more. It means we can pay for the podcast and some tasty beef ribs at the end of the year. Also, I’m always curious to know more about listeners.
Do you listen to my podcast, Software Defined Talk? If you do, I’d appreciate you taking the time to fill out our listener survey. We do occasional ads and paid interviews, and this helps us do more and charge more. It means we can pay for the podcast and some tasty beef ribs at the end of the year. Also, I’m always curious to know more about listeners. It should just take a couple minutes, I’d appreciate it if you made the time to answer the survey. Also, stickers are always free, so if you fill it out and want a reward, just email stickers@softwaredefinedtalk.com and we’ll mail them anywhere in the world, for free.
Is Your AI Assistant Creating a Recursive Security Loop? from Camille Crowell-Lee
AI-assisted coding is starting to eat its own tail: the same LLMs that write code are increasingly asked to review it, explain security decisions, and even override their own warnings. That creates recursive trust loops where “explain your reasoning” becomes an attack surface, and models can literally talk themselves out of being secure. The fix isn’t better prompts, it’s old-school architecture - separation of concerns, non-AI enforcement, and treating LLMs as assistants, not authorities.
Check out more in her article.
Perhaps the attraction is that film noir is a way of traveling into the past—the world right before I was born. When watching old films, I like to remind myself that to the people in the film, their world was just as rich as ours and felt just as “now” as ours does. Here’s John Koenig talking about the people in old photographs:
Of course, to them, it wasn’t all flickering silence and grainy black and white. They saw vivid color rushing by in three dimensions, heard voices in deafening stereo, confronted smells they couldn’t escape. For them, nothing was ever simple. None of them knew for sure what this era meant, or that it was even an era to begin with. At the time, their world was real. Nothing was finished, and nothing guaranteed.
That world is now gone. If the past is a foreign country, we’re only tourists. We can’t expect to understand the locals or why they do what they do.
I doubt that today’s young can understand the 1970s. If you cannot imagine a professor smoking in class, or the person next to you on the airplane smoking, without it seeming weird or annoying, then you’ll never understand what life felt like in the 1970s. Just as the people of 2075 will never understand that it 2025 it felt perfectly normal to walk into a friend’s house without taking off your shoes, sit down at a table, and begin eating an animal.
From Scott Sumner.