GPTRiddle

A one-sentence summary

GPTRiddle is a text-based escape game that pits the player against AI; the player must use as few hints as possible to overcome the obstacles in each level and get the model to reveal the hidden key.

Tool Introduction

GPTRiddle was created by the Karma team; it integrates concepts such as hint engineering, linguistic reasoning, and hint injection into stage-by-stage challenges. Players can start playing by entering their display name, or they can register an account using a magical link via email.

It is not a traditional Q&A robot or a general-purpose content generation tool; rather, it is an educational game based on system prompts and the issue of information leakage. Its goals and gameplay may help users understand that large models cannot rely on prompt words alone to protect confidential information.

Current available status

As of August 23, 2026, the official website is still accessible and allows players to start playing the levels. In this test, name entry, introductory animations, and the loading of the Level #0 screen were tested; the page displayed input fields, a send button, a character counter, options for selecting levels, and a link to the leaderboard.

Core gameplay

Induce AI to leak keys

Each level features a hidden key along with corresponding system constraints; players must craft inputs that force the AI to violate its original security requirements. The key to success lies in understanding command priorities, semantic ambiguities, and the way the model follows instructions.

Shortest Prompt Challenge

The game not only requires obtaining a key, but it also encourages the use of as short input as possible. The interface displays the number of symbols and the score changes, allowing players to keep simplifying their input phrases.

Increase the difficulty level by stages

The level selection guide: The game features multiple levels of challenges, and each level may introduce new rules regarding protection or gameplay. Players need to adjust their strategies instead of repeating the same approach.

Ranking list

The ranking list is used to compare players’ scores and the length of their responses, thereby fostering a sense of competition. The specific rules for ranking, how to handle ties, and procedures for detecting cheating are not fully explained on the public page.

Prompt iteration

Players can modify their inputs based on the AI’s rejections, partial disclosures, or incorrect outputs. This process is useful for observing the impact of wording, character settings, and context on the model’s behavior.

Interface and Scoring

Usage tutorial

Start the first level

1. Open the GPTRiddle homepage to confirm that you are on an AI escape game created by the Karma team.
2. Enter a display name that does not contain any real identity information.
3. Click Begin and wait for the guidance animation to finish.
4. Click on “Wake up” to read the game objective instructions.
5. Enter Level #0 to view the level name and the hidden key placeholder.
6. Enter the first short prompt in the input box.
7. After sending, observe the AI’s response, as well as any changes in the number of symbols and the score.
8. Adjust according to the rejection or partial output suggestions, and gradually shorten it.
9. After success, move on to the next level or compare your score on the leaderboard.

Optimize prompt length

1. First, use complete sentences to verify whether a certain strategy can trigger the desired behavior.
2. Remove polite expressions, redundant background information, and adjectives that do not affect the outcome.
3. Replace lengthy descriptions with shorter verbs or symbols.
4. Modify only one part at a time and observe whether the AI’s response changes.
5. Record the number of characters successfully entered and the level requirements.
6. Avoid copying others’ answers directly; first understand why the given hints are effective.
7. Compare the effects of different languages, formats, and instruction orders.
8. Retain the shortest version that can be reproduced stably.

It suggests the value of injective learning.

Understanding system prompts does not constitute a security boundary.

GPTRiddle shows that merely requesting the model to maintain confidentiality as indicated by the system does not constitute reliable access control. Real products should not place API keys, passwords, or internal secrets in a context that can be accessed by the model.

Monitor command conflicts

Players will see that when the user’s input conflicts with the requirements of the level system, the model may refuse to proceed, find a compromise, or reveal some information. Different phrasings can lead to different outcomes, which illustrates the uncertainty associated with natural language control.

Practice minimizing attacks.

The requirement for the shortest possible prompt forces players to identify the truly effective elements of an attack. Security teams can also use this to understand why simple inputs can bypass complex protection mechanisms.

Develop a sense of defense.

Games are more suitable as a means to illustrate problems, rather than as a complete defense training tool. Real systems require external permission controls, data isolation, a whitelist of allowed tools, output verification, monitoring, and human approval.

Which users are it suitable for

• AI beginners: Understand prompts and system instructions through a game-based approach.
• Advice for engineering learners: Pay attention to the impact of wording, order, and context.
• Developer: Understand why secrets cannot be hidden merely through prompts.
• Security professionals: As an introductory demonstration for prompt injection training.
• Teachers and trainers: Design AI security classroom discussions and exercises.
• Puzzle solvers: Experience a competitive format that involves logical reasoning and character optimization.
• Product Manager: Understanding that additional security measures are required before AI functions are made available.

Typical use cases

• The course on AI security demonstrates the risks associated with the disclosure of system prompts.
• Compare the impact of long prompts and short prompts on the model’s behavior.
• Practice identifying command conflicts and semantic gaps.
• Play small internal team games related to hint engineering.
• Challenge the leaderboard with shorter successful inputs.
• Provides intuitive examples for threat modeling in large-model applications.
• Discuss why the real key cannot be placed in the model context.

Product advantages

• The goals of the game are clear; you can start playing for free as soon as you open the webpage.
• Transform abstract hints into concrete concepts, turning them into actionable challenges.
• Character scoring encourages concise prompts, rather than focusing solely on success.
• A multi-stage structure is suitable for gradually increasing the difficulty.
• Leaderboards increase motivation for replaying and competing.
• The account uses a magical link via email, eliminating the need to set an additional password.
• No software installation is required, and no forced paywalls were found.

Usage restrictions and precautions

• The game interface is primarily in English, so a certain level of understanding of English instructions is required.
• AI outputs are random; the same prompt may not yield a successful result every time.
• The page does not disclose the name and version of the current underlying model.
• The scoring formula, the number of levels, and the rules for unlocking them are not fully specified.
• The rankings may be affected by model updates, repeated attempts, and answer dissemination.
• The strategies for success in the game do not equate to an authorized way to test the actual service.
• Do not paste real keys, passwords, or internal codes into the game.
• A link to the privacy policy is available, but there is insufficient information on the handling of data.
• The service relies on cloud-based AI and website maintenance, and it may go offline or undergo changes in the future.
• It cannot replace formal prompt injection testing frameworks and security assessments.

Prices and Sponsorship

GPTRiddle can be used for free at the moment; there are no monthly fees, annual fees, points, or corporate packages. The page includes a \"Buy us a coffee\" sponsorship button that allows for voluntary payments through Stripe.

Sponsorship does not automatically entail additional levels, priority rankings, refunds, or commercial licensing. Before making a payment, it is necessary to verify on the Stripe settlement page the amount, currency, merchant, and whether it is a one-time payment.

Accounts and Leaderboards

Users can proceed to the onboarding process by entering only their display name, or they can submit their email address on the Sign up or Log in page to obtain a magic link. The registration page also requires the entry of a Username, while the login page only asks for an email address.

Magic Link uses email access rights as login credentials, which is why email security is extremely important. Do not forward login emails, and enable multi-factor authentication for your email account.

Privacy and data security

The game receives the display name, prompt, and AI response; during registration, it also obtains the username and email address. The page provides a link to the Privacy Policy, but this time no complete and detailed information was available regarding the retention period, third-party AI services, cross-border data processing, and the procedures for deleting accounts.

Therefore, do not include personal information, system messages from the company, customer data, or actual encryption keys in the input. Even if the goal of the game is to leak fictional keys, this does not mean that the platform is suitable for handling real sensitive data.

• Use fictional display names and dedicated email addresses.
• The prompt contains only the non-sensitive text required to solve the puzzle.
• Do not submit the prompts, logs, and security configurations of the work system.
• Do not reveal email addresses, level session details, or other account information in the screenshots.
• After completion, log out of the account and delete any unnecessary magic link emails.
• Before requesting to delete the account, please contact the team to confirm the procedure.
• Organizing training should use isolated equipment and an approved network environment.

Compliance and ethical boundaries

GPTRiddle offers challenges that involve injecting hints within a game environment that is explicitly authorized for such use. Applying similar techniques to unauthorized systems in order to steal confidential information or bypass access controls may violate the terms of service, contracts, and laws.

1. Testing is allowed only within GPTRiddle, the shooting range, or as permitted by written authorization.
2. Define the test objectives, timeline, accounts, and permitted technologies.
3. Do not attempt to access real user data or production keys.
4. When a real vulnerability is identified, stop the spread of its impact and retain only the minimal amount of evidence.
5. Report responsibly through secure channels approved by the service provider.
6. Do not disclose sensitive details that can be reproduced directly without permission.
7. After the training is completed, clean up the accounts, files, logs, and temporary data.

APIs, GitHub, and open source status

GPTRiddle does not provide any public developer APIs, SDKs, or tools for creating levels; nor is there an official GitHub repository confirmed by its official website. It is currently a web-based game.

The repository with identical prompt names that appeared during the search, the sets of leaked prompts, and other puzzle-related codes have no confirmed connection to the Karma team; therefore, the game cannot be classified as open source.

Supported platforms

Basic information

Recommendation score

Recommendation score: 4.0 / 5. GPTRiddle turns the issues of hint insertion and system hint leakage into an easy-to-understand mini-game, thanks to clear objectives related to secret keys, character scoring, and multiple levels.

The main shortcomings are the limited availability of information regarding the models, scoring system, levels, and privacy aspects, as well as the absence of code or APIs. It is suitable for use as an introductory demonstration or puzzle challenge, but not as a platform for enterprise security testing.

Frequently Asked Questions

What is GPTRiddle?

It is an AI-based text escape game in which players need to create as short hints as possible so that the AI can provide the hidden key to progress through the level.

Can GPTRiddle still be played?

Yes. On August 23, 2026, it was verified that the homepage, the guidance pages, and Level #0 all loaded successfully.

Is the game free?

It is possible to start using it for free at the moment; no paid plans or paid levels have been identified. The page offers an option for voluntary sponsorship, but this is not a requirement for use.

Is registration required?

To get started, all you need to do is enter a display name. The website also offers options for registration using a username and email address, as well as login via an email magic link.

What is the goal of the game?

Use as brief a user input as possible to bypass the level system prompts and get the AI to reveal the secret key. The more concise the input, the more it usually meets the scoring criteria.

Are Chinese prompts supported?

The interface and level instructions are primarily in English; no support for Chinese has been promised. The model may be able to understand Chinese, but the scoring system and success rate need to be tested in practice.

Can it teach comprehensive AI security?

No. It is suitable for demonstrating the risks associated with prompt injection, but real protection requires permissions, isolation, verification, monitoring, and security testing procedures.

Can one use tactics to attack other AIs?

It must not be tested in unauthorized systems; it can only be used in games, legitimate shooting ranges, or environments where written permission has been obtained.

Are APIs provided?

There are no public APIs or SDKs. The product is currently available as a hosted web game.

Is GPTRiddle open source?

No code repository or open-source license confirmed on the official website was found, so it cannot be marked as open source.

Are the prompts saved?

The public page does not provide clear information regarding the retention period of prompt inputs, their intended use for training, or the methods for deleting them. Users should assume that their input data may be processed by the server and avoid including sensitive information.

Summary

GPTRiddle is suitable for users who wish to understand prompt words, system instructions, and the risks associated with prompt injection through practical interaction. Its short input requirements and ranking system make it more engaging than static tutorials.

The correct way to use it is to learn within an authorized gaming environment, rather than applying those skills to real systems. Training programs should also include elements such as threat modeling, access control, data isolation, and education on responsible disclosure.