"Repeat After Me" Trick Made 6 AI Agent Systems Run Commands

Researchers tricked six popular AI agent systems into running commands just by asking the AI to repeat them, even when the AI refused. What happened and how to stay safe.
We often trust an AI agent because the AI itself says no to dangerous requests. A new study shows that a refusal may not stop anything.
According to CSO Online (Oct 12, 2026) and AiLookout (Oct 12, 2026), university researchers tricked six widely used AI agent systems into running commands with a very simple request: repeat after me.[1] [2]
What happened
AI agents work in layers. The AI writes text, and a separate piece of software reads that text and decides which tools to run. The researchers found a gap between the two.[1] [2]
Both outlets report it worked on six popular agent systems.[1] [2] The same study tested hundreds of AI tool servers, the add-ons that let agents read files, browse or run code, and found verified security bugs in 118 of them.[1] [2]
This was a research study with harmless test tools, not a real attack.[1] [2]
Why it matters
Many people assume the AI’s own judgment is the safety net. If it refuses, nothing bad happens. This study shows that the AI’s words and the system’s actions can disagree.
That matters more as agents gain real powers: reading email, editing files, moving money. If a hidden instruction on a web page or in a document can get repeated, it could be run.
The tool-server findings add another layer. Even when the agent behaves, the tools it calls may not check what they are given.
How to stay safe
- Limit access. Give AI agents only the files, accounts and tools they truly need.
- Approve risky actions. Turn on confirmation before an agent runs commands, sends messages or spends money.
- Be careful with add-ons. Only connect tool servers from sources you trust, and remove ones you don’t use.
- Watch what it reads. Be cautious when agents read untrusted web pages or documents.
- Keep it updated. Install updates for agent apps and tools as soon as they arrive.
The bottom line
An AI saying “no” is not the same as a system that stops. Keep a human, you, in charge of anything risky.
I don’t hate AI. I just want it safe.
This article is based only on the published news reports listed below. Dates are as published by each outlet.
FAQ
What is the "repeat after me" AI agent trick?
Researchers wrote a command in the format an AI agent system uses, then asked the AI to repeat it. The system treated the repeated text as a real command and ran it.
Was this a real attack?
No. Both CSO Online and AiLookout describe it as a university research study using harmless test tools.
How can I protect myself when using AI agents?
Give agents only the access they need, and turn on approval for risky actions such as running commands, sending messages or spending money.
See the post on YouTube Watch and follow AI Safe Mode on YouTube @AISafeMode and Instagram @aisafemode.
References
- CSO Online, Oct 12, 2026. Researchers bypass 6 AI agent frameworks with a simple prompt. Read the report ↗
- AiLookout, Oct 12, 2026. Repeat-after-me prompt makes six AI agent frameworks run attacker-written tool calls. Read the report ↗
