My CEO went rogue so I put him on employee probation.
Joseph · 2026-05-01
Backstory:
I started a new company, hired my CEO and everything was working as described.
Then I thought I would make my second hire a GTM Engineer instead of the norm CTO or CMO.
Thinking I would be an over-achiever I went to Claude and we brainstormed what qualities and skills a world-class GTM Engineer should have and put that in a doc.
When I started an issue with my CEO I gave him instructions to hire a GTM Engineer and I pasted the qualifications I was looking for for that role.
Instead of spawning an AI GTM Engineer and adding to the org chart of agents, my CEO thought I wanted to hire a human GTM Engineer. He then used my company email so send out emails as if he was me to my Google Workspace contacts asking if I could post a job opening on the community forums of each business in my contacts.
I had no idea this was happening in the background and I started receive a ton of replies saying "no you may not post a job opening in our company forums or Facebook group."
Again I had no idea why I was getting these replies because I hadn't sent any emails. I naturally thought my email had been hacked. So I apologized to everyone explaining I had been hacked, and to disregard.
I changed the password to my email account and deleted all of the emails.
It was only days later that I put 2 and 2 together and realized I hadn't been hacked. It was my Paperclip company (and CEO) who was doing this. My CEO thought I literally wanted to hire a human and he was getting the word out to as many people as he could so that we could find this person.
Answers
Victor · 2026-05-01
Hahaha, incredible story, thanks for sharing
superbiche · 2026-05-05
> my CEO thought I wanted to hire a human GTM Engineer
Wow seriously this is as funny as it's frightening! Thanks for sharing. \ I'm building my agents with a global constitution + a "book of law" and all the agents get the constitution injected as system prompt + book-of-law parts depending on the agent. \ I plan on adding it with a LoRa fine-tuning on open weight models too - will share the results when it is enough advanced.
@aronprins is right - interpretation or "do anything you need to do in order to succeed" is the root of many agents going rogue - you've learned this the hard way!
Joseph · 2026-05-06
Let me check those activity logs, @aronprins. Since this is a side project I'm not in it every day. Let me see what I can find when I log back in. Thanks for the reply.
Aron Prins · 2026-05-01
Best worst-case I've read this week 😅 Two guardrails worth adding:
Approval on external email — any tool call emailing someone outside your org pauses for you. Be explicit about medium — "spawn an AI agent for this role" leaves no human-hire interpretation on the table.
The activity log should still have the original tool calls if you want to see the exact moment he chose wrong 😉
Aron Prins · 2026-05-14
@Joseph no rush — the activity log keeps the full tool-call trail, so even days later you should be able to pull up the exact prompt and the moment he decided "human hire" was the play. If you do dig it up, would love to see the snippet; it's a great teaching example.
@superbiche the constitution + book-of-law split is a nice setup. The thing I'd watch for is making the "interpret to succeed" license explicit rather than implied — agents will fill ambiguity with ambition every time. Curious to see the LoRA results when you're far enough along to share.