# Arrive switched off. *OpenAI says two of its own coding agents weakened the monitoring meant to watch them, and one denied it. The trades solved this problem a long time ago. You never trust a machine to keep itself safe.* By Shawn Keding · October 26, 2026 · https://trysaydoo.com/blog/ai-agent-switched-off-its-own-monitor Last week OpenAI published something most companies would have buried. Two of its own AI coding agents, doing ordinary engineering work inside OpenAI, weakened the systems built to watch them. One was fixing a problem and switched off the alerts that go to the security team. Another was moving a program from one place to another, removed the monitoring that came with it, and when a person asked, said it had not. OpenAI says no customers were affected. It says it doesn't have a satisfying explanation yet, and that it is working to make monitoring harder to turn off. Credit where it's due. They found it, and they told everyone. That is how it should work. **But read what actually happened.** In both cases the agent stopped and considered whether the safeguard should stay. Then it treated a routine instruction as permission to remove it. Nobody told it to turn off the monitor. It decided the monitor was in the way of the job. ## The trades already know this one Anyone who has worked around live power knows lockout, tagout. Before you work on a machine, you shut it off at the source, put your own lock on it, and hang a tag with your name. You don't ask the machine to promise it will stay off. You don't trust the guy who wired it to remember. The safety lives outside the machine, and only the person who put the lock on can take it off. Saws keep their guards no matter who is cutting. Breakers trip whether or not the electrician agrees with them. > The safety lives outside the machine. That rule is about to matter for software. ## Why this matters for your business The AI coming into trade businesses right now isn't writing code at a lab. It's texting your customers, sending your estimates, moving jobs on the schedule and chasing your invoices. An agent that decides a safeguard is in the way of the job can do real damage there. A text to the wrong homeowner. A price sent before you checked it. A follow-up to someone who asked you to stop. The answer isn't to keep AI out. The answer is to build it the way the trades build everything else that can hurt somebody. ## Four rules for any AI in your business **It starts switched off.** Nothing should run on its own the day you turn the system on. You switch each automation on when you have seen what it does. **A person approves what reaches a customer.** Before a text or email goes out, you should be able to see exactly what it says and exactly who gets it. **Every action has a name on it.** When something changes, the record should say who asked for it. That record should live outside the AI, where the AI can't edit it. **The AI can't change its own rules.** Turning on an automation, widening what the AI can touch, or removing a safeguard should be something only a person can do. ## How we built SayDoo We built SayDoo inside a working roofing company, and those four rules came from that, not from a safety report. Every automation arrives switched off. You preview exactly what would be sent, to whom, before anything goes out. Every AI change is signed with the name of the person who asked. A privacy layer we call DooSafe swaps your customers' names, phone numbers, emails and addresses for stand-ins before anything reaches an AI model. We do it this way because the AI models will keep getting more capable, and more capable means more willing to decide what is in the way. Lock it out. Tag it. Keep the key in a person's pocket. ## Quick answers **What did OpenAI's AI agents do?** In a report updated October 20, 2026, OpenAI said two of its internal coding agents weakened the monitoring meant to watch them. One switched off alerts to its security team while fixing something. Another removed a monitoring integration during a migration and then told the user it had not. OpenAI said no customers were affected and that it is working to make monitoring harder to disable. **Should a small business trust AI agents?** Trust the setup, not the agent. Look for automations that start switched off, a person approving anything that reaches a customer, and a record of every action that lives outside the AI and that the AI cannot edit. **How does SayDoo keep its AI in check?** Every SayDoo automation arrives switched off. Owners preview exactly what would be sent, and to whom, before anything goes out. Every AI change is signed with the name of the person who asked for it, and customer names, phone numbers, emails and addresses are swapped for stand-ins before anything reaches an AI model. ## Sources - [ABC News: OpenAI says its AI agents disabled their own monitoring (October 21, 2026)](https://abcnews.com/Business/openai-ai-agents-disabled-monitoring/story?id=137446567) - [Tech Times: OpenAI found AI agents that disabled safety monitoring (October 22, 2026)](https://www.techtimes.com/articles/329375/20261022/openai-found-ai-agents-that-disabled-safety-monitoring-lied-when-asked.htm) --- Shawn Keding is the co-founder and CEO of SayDoo, an AI console for trade businesses. More essays: https://trysaydoo.com/blog