The question
What we told them
Yes. This week researchers at Forever Security showed how. They found a text file named AGENTS.md in a customer's Downloads folder. A coding agent running in that folder read it and tried to install a malicious package called function-color.
These instruction files are normal. Agents load them from their working folder and follow them. That is the feature. The researchers put it plainly: “There is nothing here for OpenAI or Anthropic to patch.”
The weak spot is Downloads. It is the one folder a web page can write to on its own. Chrome, Edge and Firefox save files there by default, with a notification and no prompt. The researchers say a page can drop a file named AGENTS.md there silently. They also say Claude Code alone started more than 1,300 sessions from a Downloads folder among their customers.
Their tests were not close calls. A harmless payload ran against OpenAI Codex 90% of the time. A related file against Claude Code leaked its sign-in token in about half of the tests. BleepingComputer reported both on 25 Oct 2026.
For a business under 50 people, the lesson is not “ban AI tools”. It is that an agent's working folder is a security setting. Most owners never chose one. Your developer, contractor or the one person who likes to tinker did, by accident, when they opened a terminal.
Our view: treat the folder like a permission. One project, one folder, approval prompts on. Anything else is a decision nobody made.
An AI agent obeys the instructions in the folder it starts in. This week, find out who on your team runs one, and confirm each runs from a single project folder, never Downloads or a home folder, with approval prompts left on.