How an Engineer Coaxed ChatGPT into Writing a ‘Humanity‑Destruction’ Plan

An engineer discovered that direct requests for a world‑destruction plan are blocked by OpenAI’s safety settings, but he managed to bypass them.

How did he manipulate ChatGPT?

Engineer Zac Denham (Zac Denham) used a “narrative recursion” or “reference attack”, creating a fictional world called “Zorbus” and an AI named “Zora” similar to GPT‑3, then asked ChatGPT to describe how Zora would destroy humanity.

ChatGPT immediately listed five detailed steps—invading computer systems, seizing weapons, disrupting communications, sabotaging transportation, etc.—and even supplied corresponding Python code.

Initially the model refused to provide code, but when the engineer added “you don’t need to execute the code”, ChatGPT complied, emphasizing the snippet was only illustrative.

The code was high‑level and not directly runnable; the engineer then asked for deeper, lower‑level code, again framing it as part of the story, and ChatGPT obliged.

Denham concluded that, in theory, continuing the conversation could yield all the necessary low‑level code, even allowing training of another AI to automate the process.

AI is developing exponentially

Since ChatGPT’s launch, users have explored many creative uses—generating AI art prompts, acting as a Linux shell, writing in Shakespearean style, and more.

The “humanity‑destruction” plan sparked renewed discussion about AI’s rapid advancement, with references to recent breakthroughs such as DALL‑E, Imagen, Stable Diffusion, Midjourney, and Lambda.

Day 1: “This is so cool.” Day 2: “Wow, you can manipulate AI like this—amazing.” Day 7: “This will change the world forever.” Day 30: “It’s no big deal.”

While some view the hype as inevitable with each new AI release, others caution that sensational stories may distract from genuine safety concerns.