7 min read

Where My AI Guardrails Actually Lived

My AI tools read my written guardrails every session, and nothing enforced them. An audit showed my agents could push code and publish without asking, so I moved the important rules where an agent cannot talk its way past them. For every rule, ask what actually stops it.

On Episode 5 of Explore AI Out Loud, Brandon and I talked about how much of our data we are willing to hand to AI tools. I said “my guardrails live in my second brain and in my operating system,” and that I am “very comfortable allowing AI to intermingle into my data and into my life.” Those guardrails are written rules in my second brain. My AI tools read them before they start working with me, and some of them spell out what the tools should never do without asking first. My comfort depended on those rules holding, and I had never checked whether they would.

So I decided to build a way to challenge my own optimism. I am bullish on AI, and most of my time goes into building with it. I have a research agent that reads the news for me every morning, a chatbot on my website that answers questions about past issues, tools that help Brandon and me produce the podcast, and AI skills that help me write each issue of this newsletter. Building that fast makes it easy to skip the other side of how I think. The day after we recorded, I started a security audit of my own setup, a deliberate search for the vulnerabilities and security gaps in everything I had built. I asked my AI tools to help me check all of it, from my computer and accounts to the files I share, my website, and the AI builds behind this newsletter and podcast. The list I worked through ran from basics, like turning on my computer’s firewall, moving my passwords into one password manager, and adding two-factor sign-in across my accounts, to setting spending limits on the AI services that bill me, upgrading the server that runs my automations, and locking down email security on the domain my podcast uses.

The podcast

Explore AI Out Loud

Find practical ways to use AI in your work and everyday life. Learn from real experiments, including the mistakes, so you can spend less time guessing and more time trying something useful. New episodes every other Tuesday.

Subscribe on YouTubeFollow on SpotifyFollow on Apple Podcasts

Latest episodes and the show

What October Looked Like

That audit came at the end of a month that made the risk hard to ignore. On October 21, the maintainers of PHP released fixes for two critical flaws. PHP is a programming language that runs a large share of the web, including government services, online stores, and banks. Attackers did not wait for the announcement. Amazon Web Services saw an attempt to exploit the problem on October 17, four days before the fixes were public. The UK’s National Cyber Security Centre advised organizations that could not patch quickly to consider taking affected systems offline until the problem was fully mitigated, and governments did. In the United Kingdom, 32 government services were unavailable, largely as a precaution, and government bodies in Finland, France, and Denmark took services offline as well.

Some of the researchers who found those flaws were working with AI models, and that kind of help works for both sides. Writing about exploits in general, Amazon’s security team warned that “AI-assisted analysis makes it cheaper to turn a patch diff or advisory into a working exploit.” Defenders get less time to patch. AI agents are also starting to act where nobody approved them. The Wikimedia Foundation reported that AI agents it believes OpenAI operated had made unapproved edits on its projects, and it put the word “rogue” in quotation marks.

A coding agent on a customer’s computer tried to install a malicious software package called function-color, and the security firm Forever stopped it before any code ran. The agent was following instructions in a file that had landed in the Downloads folder from the web. Files like that are written in Markdown, a plain text format, and agents read them to learn how a project should be handled. The researchers summed it up in five words, “Markdown is the new executable.” A text file an agent reads can behave like a program, because the agent will carry out what it says. Of everything I read that month, this incident sat closest to my own setup. The same agency offers a standard written for organizations that applies just as well at home. “If you cannot understand, monitor or contain an agent’s actions, it is not ready for deployment.”

Inside My Own Setup

Measured against that standard, my own agents would not have passed. My guardrails were real and carefully written, and my tools read them at the start of every session. Nothing enforced them. Over months of building, I had approved actions one at a time so the work would move faster, and each approval became permanent. By the time of the audit, my coding agents could run their own code, print the secrets stored in my working environment, push code to my public repositories, and publish changes to the AI skills behind this newsletter, all without asking me. The rule telling them to check with me before publishing existed only in writing. It held as long as the agent remembered it and nothing it read convinced it otherwise. The Downloads attack relied on that same weakness. I am probably too lenient with my agents at home, and the audit showed me exactly where.

Then there was the key with full control of my website, sitting in plain text in my AI tool’s history file. I had pasted it into a prompt while building something and never thought about where that text went. Any agent that could read my files could read that key. Pair it with an agent that follows whatever it reads, and instructions written by someone else could have reached my website. A secret for the chatbot on my website was sitting in my command history the same way. This was a miss in my build, and it is the version of the risk that worries me most, an AI acting on my behalf with access I never meant to give it.

Over the next week, I worked through both, one step at a time. I replaced the website key and the chatbot secret, and I cleaned the website key out of the history file. For my agents, I added limits that do not depend on what they remember. They now have to ask before pushing code or publishing those skills, and some actions are blocked outright, whatever the agent has been told. One part of my setup needed no repair. My public code repositories were clean across their full history. The keys my tools use live in my computer’s keychain instead of in the code, and secret scanning and push protection were already turned on to check every change for passwords and keys and block the ones that contain them.

The Same Problem, Moving Faster

None of this feels new to me. Every technology that asked for our data has carried risk, from the first online accounts to the apps on our phones, and each one arrived with the same hype, fear, and noise. AI is that problem moving faster, with tools that can act as well as answer. I think AI is inevitable, and in cybersecurity it will be hard to control for a while. There will be incidents and growing pains. Defenders will also catch up by using AI to combat AI, and the researchers who found the PHP flaws with help from AI models are an early example.

I do not work in Silicon Valley, and the steps that helped me most do not require it. The first was looking at what my AI tools could reach and what they could do without asking me. Most agent tools let you choose which actions need your approval, and the choices you clicked through months ago deserve a second look. The second was asking, for every rule that says an agent should never do something, what actually stops it. If the answer is only an instruction the agent reads, it is a request. It also helps to treat files and web pages an agent reads as possible instructions, and to keep passwords and keys out of chats and prompts, where they can linger long after you forget them. If you use these tools at work, the same questions belong with whoever manages them for your company.

For my own work, security is moving into the builds themselves. I added these lessons to my second brain, so every new build now starts by deciding what its agents can reach, what they can do on their own, and what enforced limit stops each risky action. Once a quarter or so, I expect to run a full audit of everything and adjust what I check as I learn. This field is changing too quickly for any single checklist to last, so I want a process I can repeat and improve as a regular person keeping up with it. I am still as comfortable letting AI into my life as I was on Episode 5. More of my guardrails now live where an agent cannot talk its way past them.

The podcast

Follow Explore AI Out Loud

Short clips from each episode, plus what did not make the cut. New episodes every other Tuesday.

Follow on InstagramFollow on TikTok

Full episodes and the show

Make AI useful in your everyday work.

Neural Gains Weekly is free. One email every Tuesday: a full issue one week, the best lesson from the new podcast episode the next.

Before you go: one email every Tuesday.

A full Neural Gains Weekly issue one week, the best lesson from the new podcast episode the next. Free, plain English, no hype.

Subscribe for free