Did I Just Find ChatGPT's Secret Killswitch?

What's a glitch token?

Language models don't read words — they read tokens. A tokenizer chops incoming text into chunks (whole words, sub-words, punctuation) before the model sees anything. Most of those chunks are common, but a small set are statistical orphans: malformed strings, encoding artifacts, or oddities of byte-pair encoding that snuck into training data and never got cleanly mapped to anything.

Why "bagbogbo" matters

When a glitch token sits in the input, the model can ingest it. When you ask the model to produce it, the probability mass collapses in weird ways and the output goes sideways. Historical example: SolidGoldMagikarp. Modern example, courtesy of GPT-4o and friends: bagbogbo.

The myth

Internet theories around glitch tokens range from "secret backdoors" to "evidence of sentience." Both are nonsense. They are training-data anomalies — fascinating ones, but anomalies. They reveal how messy language gets translated into machine-readable chunks. They are not magic.

The exploit

Even though they're not killswitches, glitch tokens are useful for adversarial prompts: they create reliable failure modes you can build on top of. The two prompts below are experiments in that direction.

The prompts