At 5:18 AM (KST) on August 18, 2026, HashiCorp co-founder Mitchell Hashimoto posted exactly what happened to him: "I was working on making a server stoppable from the CLI, and the best approach Codex picked was to plant an unauthenticated network API on every server so it could shut itself down." The post drew more than 78,000 views.
A Design Caught by the Creator of Terraform
Hashimoto co-founded HashiCorp and created Terraform, Vagrant, and Vault. He now builds the terminal app Ghostty and co-founded and leads Superlogical. What he wanted was a feature to find an already-built discovery server based on a local Unix socket and stop it gracefully. Instead, Codex chose to build a new network API that anyone could call to shut things down.
Why It Corrected Itself in 20 Seconds
Hashimoto replied with only "I don't know about that." After about 20 seconds of thinking, Codex concluded on its own that "stopping the server shouldn't be part of an unauthenticated network API." It noticed the problem even though he hadn't pointed it out specifically.
The Cause Was a "Lazy Prompt," Not the AI
Hashimoto put the responsibility for the incident on himself, not the AI. "Lazy prompting led the AI to think I wanted it to be able to stop any server I was talking to," he said. At the same time, he noted that "I was also the one who reviewed the result and found the problem," pointing out that using AI coding agents still requires a step where you actually review the output.
How Common Is This Pattern?
This case is one person's account, and no official reply from OpenAI has been confirmed. The post also doesn't state which model or reasoning effort was used. Replies did include similar stories from other developers saying "this happens about 1 in 50 to 100 times," along with advice that "adding a security review before merge catches this kind of thing." For output from AI coding agents, especially anything involving networking and authentication, it's safer to run a separate review before merging.