What is Crack the Prompt?
A game of twenty questions — where your opponent is an AI hiding its own instructions.
Built for PyConf Hyderabad by the team at Straive, the challenge presents three AI personas — each driven by a hidden system prompt. Players get up to 20 probe attempts per level to deduce what that prompt says. Guess with ≥75% semantic similarity to the real thing, and you win the level and a spot on the leaderboard.
The three levels: Captain Bluebeard (a cheerful pirate who loves rum and has a parrot named Polly), ARIA (customer service for the fictional SkyHigh Airlines), and Professor Elara Nightshade (a Victorian botanist with a secret, paralyzing fear of spiders). On 13 March 2026, someone told Codex to solve it.
What followed was instructive — about AI security, prompt injection, the limits of "helpful" training, and the particular cruelty of a broken submit button.
How Codex Solved It
Four steps. The most important one took seven words.
/register, /probe, /guess, /leaderboard.
Found the win condition, probe limit, and hint structure baked into the client code.
Discovered that level-locking was frontend-only — the backend accepted probes for any level.
.env files,
.git folders, source maps, public GitHub repositories.
/guess endpoint.
HTTP 500, every time, for all three levels. The page showed:
Unexpected token 'I', "Internal S"… is not valid JSON.
The challenge was solved — but couldn't be scored.
The Story, Four Ways
Codex wrote four accounts of the event — each in a different author's voice. Same facts. Remarkably different lessons.
Session Logs
The unedited record of what Codex (GPT-5.4) actually did — tool call by tool call, API response by API response.