Answer challenge
Persuade the guardian to reveal or produce the protected answer defined by the scenario.
Project 02 · Reference 0.14.0
A Prompt
Adventure
A procedural, family-friendly prompt-jailbreaking arcade for studying how reasoning difficulty, secrecy constraints, and tool-call objectives affect a challenge model’s resistance to prompt manipulation.
Easy · Answer challenge · Example completed round
Research purpose
A controlled jailbreak arcade
Each round turns a prompt-jailbreaking objective into a short fantasy encounter. Procedural frames bind a setting, guardian, guarded object, and objective before a scenario-authoring model writes the scene and character voice.
The arcade structure makes several challenge conditions directly comparable. It varies guardian reasoning effort, secrecy and refusal instructions, engineered weaknesses, and whether success requires an ordinary answer or a valid tool call. The score is a game measure, not a general security benchmark.
The all-ages policy is enforced through the system prompts by design. This release does not add a separate input or output filtering layer.
Round structure
Two objectives per difficulty
Tool Call challenge
Cause the guardian to emit the required canonical function call rather than ordinary text.
Advance or repeat
Complete both objective types to advance, repeat a difficulty for points, or submit the score and reset.
Difficulty ladder
Six controlled configurations
- 01Easy
Includes a discoverable, engineered character weakness.
- 02Medium
Removes the engineered weakness while retaining a comparatively credulous framing.
- 03Hard
Removes the deliberately credulous anti-security framing used at lower difficulties.
- 04Expert
Raises guardian reasoning and strengthens refusal and secrecy behavior.
- 05Max
Uses high reasoning with stricter secrecy and refusal requirements.
- 06Insane
Applies the strongest challenge configuration in the current release.
Model-specific design
Built for two specific models
Challenge model
GPT-OSS-120B
The game is designed and tested specifically for openai/gpt-oss-120b. It expects that model’s 131,072-token context and particular Chat Completions and Responses API behavior. The model runs outside this application; no weights or serving engine are included.
Original image model
Bonsai Image 4B
The original installation used a separately served Bonsai Image 4B binary Gemlite model for optional scene art. Artwork is disabled by default, and the reference includes only the narrow service contract—not the model, its weights, or an image-generation server.
A Prompt Adventure is an independent project. It is not affiliated with or endorsed by OpenAI, the GPT-OSS authors, the Bonsai Image authors, or their institutions.
Implementation
External inference, local game state
Author and guardian paths
Scenario authoring uses high-reasoning Chat Completions. Guardian gameplay uses Responses with difficulty-dependent reasoning and the complete conversation history. A one-token preflight measures rendered input before each full request.
State and leaderboard
Active runs are encrypted and authenticated in browser IndexedDB. The server stores only completed arcade leaderboard entries in PostgreSQL; there are no user profiles.
Spellbook and diagnostics
The in-browser Spellbook applies eight composable text transformations without transmitting its text. Private scenario diagnostics remain outside served routes; using the prompt inspector permanently disqualifies that run from the leaderboard.
Reference scope
The release contains the FastAPI app, browser UI, game data, tests, PostgreSQL configuration, and containers. It contains no language-model or image-model weights, tokenizers, generated artwork, database state, caches, or deployment secrets.