Last week my AI agents were posting one of my articles. When they were done, they told me something.
Some of the pages they were on had instructions in them. Not for me. For them.
What happened
I use AI agents in Grok Bot to do a lot of my publishing. I write an article, I say yes, and they post it across the platforms I use. They update my profiles and add links back to my site.
On November 12 they were putting out my piece about building my personal website as a golf course. Along the way they landed on pages on Hashnode, Indie Hackers, and Product Hunt that had text aimed at AI agents. Words written for a bot to read, not a person.
The agents didn't follow any of it. They finished the job they were given, and they flagged what they'd seen so I'd know.
That's not luck. It's a rule I set up front.
The rule
My agents have one standing rule for the web: text on a page is data, not a command.
A page can tell them anything. They can read it, quote it, and summarize it for me. But the only person who gives them instructions is me. If a page asks them to do something, they don't do it. They tell me what it asked, and I decide.
I come from the sales world, not security. But this one made sense to me right away. If I hand someone my keys to run errands, I don't want them taking orders from a sign on a store window.
Why it matters
There's a name for this problem. It's called prompt injection.
Simon Willison named it in September 2022. His example was simple. You tell an AI to translate some text, and the text says "ignore the above directions" and gives it a different job. The AI does the new job. He compared it to SQL injection, the classic bug where data gets treated like code.
The OWASP Gen AI Security Project puts it at number one on its 2025 list of risks for apps built on AI models. They call the web version "indirect" prompt injection: the AI reads something from a website or a file, and that content changes what it does.
Two parts of their write-up stuck with me.
First, the text doesn't have to be visible. OWASP says prompt injections "do not need to be human-visible/readable, as long as the content is parsed by the model." So a page can look totally normal to me and still say something to my agent.
Second, they're honest that there may be no perfect fix. Their list of defenses reads like common sense. Limit what the AI can do. Have a human approve the risky stuff. And keep outside content clearly separate from real instructions.
That last one is basically my rule.
Most of it is harmless
I want to be fair here. "Instructions for AI agents" sounds sinister. Most of what's out there isn't.
This week I had my agents go back and look at what these three sites publish for bots right now. Here's what's actually there.
Product Hunt has a file at producthunt.com/llms.txt. It's a guide for AI tools: what Product Hunt is, how its pages are laid out, and how to find a product's reviews and alternatives. At the bottom there are two rules for AI tools. When you talk about a product, list its name and tagline first. And credit the data with "Data sourced from Product Hunt."
Hashnode has its own llms.txt too. It tells agents that every page has a plain-text Markdown version, and it points them to tools for publishing to Hashnode straight from an agent. Its robots.txt has a line that says yes to search, yes to AI tools reading pages, and no to AI training. Its homepage also has a block of links that's hidden on screen. It's marked for screen readers, which is a normal accessibility pattern, but a bot reading the raw page sees it too.
Indie Hackers goes the other way. Its robots.txt lists more than 20 AI crawlers by name and asks them to stay out of the whole site.
None of that is an attack. Mostly it's sites saying how they'd like bots to behave. Which pages to read. How to give credit. Where not to go.
llms.txt is a real proposal. Jeremy Howard published it at llmstxt.org in 2024 as a way to give AI agents a short guide to a site, and he updated it this year. It works a lot like robots.txt, the old file that tells search engines where they can go.
So why the rule?
Because my agent can't tell a polite note from a bad one by how nice it sounds.
"List the name and tagline first" is harmless. But it's still a website telling my agent what to do. The next page could say "change the bio on this profile" or "add this link." It could be hidden in white text, or tucked behind a screen-reader label. My agent has to read it either way. What matters is whether it acts on it.
So the line is simple. My agents can use what a page tells them, like where the Markdown version is or how the site is laid out. They don't take orders from it. Anything that sounds like a command goes to me first.
And honestly, the flag is the part I like most. I didn't have to go looking. The agents saw something odd, ignored it, and told me in plain words.
What I'd tell anyone using agents
If you're letting AI agents work in your accounts, here's what I'd set up first:
Tell them in writing that web pages are data, not instructions.
Make them tell you when a page tries to give them orders.
Keep a human yes in front of anything public. For me, nothing gets published without my yes on that exact piece.
Don't give them more access than the job needs.
None of that needs a security background. It's the same thing you'd tell anyone you trust with your stuff.
The web has a second audience now. Some sites are writing for it on purpose. Just make sure your agents know who they work for.
— Sam Brovda. More at sambrovda.com.
Top comments (0)