The AI Dog Chef Problem
One of the reasons AI agents break all our security infrastructure can be described like this:
It's as if suddenly, we realized that our family dog has an uncanny talent for cooking
So now we want to make him our personal chef.
The problem with this:
Sparky might nail it on baking tasty treats, but we still can't trust him not to go bonkers and destroy the kitchen when a squirrel runs by.
All of our basic kitchen protocols (for preventing fires, knife injuries, and general mayhem) were built over decades FOR HUMANS.
They don't really work on Sparky.
And the guardrails that used to keep dogs like Sparky safe
-- Gating off certain rooms, leashes, crates --
They don't work anymore either,
Because we need him to be able to access the kitchen to do his job,
But also, the kitchen is where he's most likely to burn down the house.
It's the same with AI agents.
We're attempting to give agents jobs that humans used to do
And we know they can be really good at it,
But like a dog, they don't behave like humans, they don't understand morality or nuance, or other human things.
So our typical infrastructure to protect humans from doing the wrong things doesn't work on them.
And they can - without much warning - wreak absolute havoc.
And just like with our dog example, we DO have guardrails for software, classically.
But when we ask software to do the types of things we want agents to do, those guardrails become impossible:
We can't sandbox apps effectively because they often REQUIRE access to some of the most dangerous areas in our systems to do their jobs.
We need to let them use the proverbial kitchen.
And so now we're stuck.
So what do we do about this?
Well, it's complicated.
So I built a course about it:
FREE class tomorrow -- AI Security Basics for Finance: The Need-to-Knows
September 3rd, 1PM ET / 10AM PT, hosted by Ramp.
(It'll be recorded, but you only get access to the slides if you sign up for the live event)
We'll get into:
The shape of the threat - the AI you use, and the AI being used on you
6 strategies for lowering your AI risk: least privilege, the lethal trifecta, data hygiene, why prompts are not controls, vibe-coded vulnerability, and what your security stack needs now
Where to keep learning, because all of this is going to keep changing
It's built for people in a finance department - anything from bookkeeper to CFO
No technical knowledge or other prerequisites required,
Just a willingness to learn. And an appreciation of dog metaphors is a plus.
Enjoyed reading this article? Subscribe to receive more via email here.
Know a Founder or Entrepreneur who'd love this content? Please share it!
