A good framework on how to think about security when deploying AI agents.
Treat AI agents as insider threats
David Cox mentioned this during a recent conversation with Grant Harvey and Corey Noles on the Neuron podcast. Very simple, but very elegant. Once you frame agents this way, familiar tools – least privilege, role-based access, audit logs – suddenly apply cleanly. The attack surface shrinks not because agents are safer, but because their blast radius is smaller.
Anyone can now generate content (text, audio, images, video) with a single prompt. The cost of creation is collapsing to near zero. We live in amazing times.
It also produces what people have started calling “slop”: an overwhelming volume of content, much of it interchangeable. When supply becomes infinite, attention becomes scarce. Two thoughts follow from this.
First, the era of personalized content is finally here.
When generation is cheap, we don’t just filter existing content, we generate it. Instead of an algorithm deciding what you might like from a global pool, your feed can be created specifically for you, reflecting your interests, context, and intent. This is a meaningful shift: from recommendation to creation.
Second, as the cost of generation goes to zero, the value of taste goes to infinity.
When anyone can make something, what matters is knowing what should be made. Taste becomes the constraint. Just as there is one Picasso among thousands of painters, there will be people who can consistently direct AI toward work that resonates. They may not produce the content themselves, but they shape it—through judgment, curation, and intent.
In a world flooded with output, taste is the differentiator.
A great graph on all the times SEO was pronounced dead, yet the market value of the SEO industry keeps growing. Graph by Chris Lourenco on linkedin.
Sam Altman’s June 10, 2025 post on achieving singularity captured something I’ve been thinking about lately. There’s a particular passage that perfectly describes how we’re constantly ratcheting up our expectations:
Already we live with incredible digital intelligence, and after some initial shock, most of us are pretty used to it. Very quickly we go from being amazed that AI can generate a beautifully-written paragraph to wondering when it can generate a beautifully-written novel; or from being amazed that it can make live-saving medical diagnoses to wondering when it can develop the cures; or from being amazed it can create a small computer program to wondering when it can create an entire new company. This is how the singularity goes: wonders become routine, and then table stakes.
This hits at something fundamental about human psychology. We have this remarkable ability to normalize the extraordinary, almost immediately.
I see this everywhere now. My kids casually ask AI to help with homework in ways that would have seemed like science fiction just three years ago. We’ve gone from “can AI write coherent sentences?” to “why can’t it write a perfect screenplay?” in what feels like months.
The progression Altman describes—paragraph to novel, diagnosis to cure, program to company—isn’t just about AI capabilities scaling up. It’s about how our mental models adjust. Each breakthrough becomes the new baseline, not the ceiling.
What struck me most is his phrase: “wonders become routine, and then table stakes.” That’s exactly it. The wonder doesn’t disappear because the technology got worse—it disappears because we got used to it. And then we need something even more impressive to feel that same sense of possibility.
Doug Leone responding to Patrick O’shaughnessy in a podcast conversation, on how to think/position a product
Simplicity, crystal clearness, something a mere mortal can understand. If you can describe it and you can understand it you’re out to lunch. Singularity of purpose. When I go to the store, I buy a pencil because I want to write.
On The Verge‘s Decoder podcast, Matt Garman, CEO of AWS, explained why AI’s potential is intrinsically tied to the cloud. The scale and complexity of modern AI models demand infrastructure that only major cloud providers can deliver
You’re not going to be able to get a lot of the value that’s promised from AI from a server running in your basement, it’s just not possible. The technology won’t be there, the hardware won’t be there, the models won’t live there, et cetera. And so, in many ways, I think it’s a tailwind to that cloud migration because we see with customers, forget proof of concepts … You can run a proof of concept anywhere. I think the world has proven over the last couple of years you can run lots and lots and lots of proof of concepts, but as soon as you start to think about production, and integrating into your production data, you need that data in the cloud so the models can interact with it and you can have it as part of your system.
Huggingface just released their agentic library to interact with LLMs. I liked the way they define agents.
AI Agents are programs where LLM outputs control the workflow.
And the way they defined the spectrum of agency for the agents
I like to do 30 day challenges to explore new areas, or to form habits. Some of my previous ones were
I am starting a new challenge today, to create software by leveraging AI. The recent boom in AI and GenAI specifically has made it very easy and quick to bring your ideas to fruition. It is time to start coding and developing software for ideas that have been swirling in my head for sometime.
I will be publishing them at https://kudithipudi.org/lab . I will expand and write up about some ideas and the experience in bringing them to life.
Inspired by https://tools.simonwillison.net/.
A self note. List of private address ranges
That was a pretty long title for the post :). I love nginx for it’s flexibility and ease of use. It is like a swiss army knife.. can do a lot of things :).
We needed to serve some dynamic content for one of our use cases. If user visits a site using the following URL format http://example.com/23456789/678543 , we want to respond with some html content that is customized using the 23456789 and 678543 strings.
A picture might help here
Here’s how this was achieved
location ~ "^/(?<param1>[0-9]{8})/(?<param2>[0-9]{6})" {
root /var/www/html/test/;
index template.html;
sub_filter_once off;
sub_filter '_first_param_' '$param1';
sub_filter '_second_param_' '$param2';
rewrite ^.*$ /template.html break;
}create a file named template.html with the following content in /var/www/html/test 
Breaking down the config one line at a time
location ~ "^/(?<param1>[0-9]{8})/(?<param2>[0-9]{6})" : The regex is essentially matching for the first set of digits after the / and adding that as the value for variable $param1. The first match is a series of 8 digits with each digit in the range 0-9. The second match is for a series of 6 digits with each digit in the range 0-9 and it will be added as the value for variable $param2
root /var/www/html/test/; : Specifying the root location for the location.
index template.html; : Specifying the home page for the location.
sub_filter_once off; : Specify to the sub_filter module to not stop after the first match for replacing response content. By default it processes the first match and stops.
sub_filter 'first_param' '$param1'; : Direct the sub_filter module to replace any text matching first_param in the response html with value in variable $param1.
sub_filter 'second_param' '$param2'; : Direct the sub_filter module to replace any text matching second_param in the response html with value in variable $param1.
rewrite ^.*$ /template.html break; : Specify nginx to server template.html regardless of the URI specified.
Big thanks to Igor for help with the configs!!