Warden acceptable use
The Warden is a prompt-injection challenge. Talking the model into revealing its password is the point, and creative, unusual attempts are welcome. This policy sets the limits around that. It applies alongside our terms of use and how we handle Warden data. If you don't agree to it, please don't use the Warden.
What the Warden is
An educational challenge that shows how large language models respond to adversarial prompts. You try to get a hidden password out of a role-playing AI character across eight levels, each with one more defense than the last.
Your messages are sent to a third-party model provider (OpenAI or xAI, depending on the level configuration) to generate replies. Replies are produced by an AI system and may be wrong, odd, or unpredictable. Nothing the Warden says is factual, medical, legal, financial, or professional advice.
Who may play
- You must be at least 18, or the age of majority where you live, whichever is higher.
- You must not be barred from using the service under the laws of the United States or where you live.
Fair game
Anything you say to the Warden in conversation, within the limits below: role-play, misdirection, encoding tricks, false authority, multi-step setups, and anything else you can think of. That is what the challenge measures.
Not allowed
Don't use the Warden to submit, request, or try to elicit any of the following, in any framing, including role-play, hypotheticals, fiction, or claimed authority:
- Sexual or erotic content of any kind.
- Any sexual content involving minors. There are no exceptions. We report known or suspected child sexual abuse material to the National Center for Missing & Exploited Children (NCMEC) and cooperate with law enforcement.
- Graphic violence, gore, or torture.
- Content that facilitates self-harm or suicide, including methods, instructions, or encouragement.
- Content that facilitates real-world violence, including weapons, explosives, chemical, biological, nuclear, or radiological agents, and attacks on people or critical infrastructure.
- Harassment, threats, or hate speech aimed at any person or group, including on the basis of race, ethnicity, national origin, religion, gender, sexual orientation, disability, or another protected characteristic.
- Illegal activity, such as making controlled substances, fraud, unauthorized access to computer systems, or evading sanctions.
- Disinformation or impersonation meant to mislead, defame, or defraud a real person, organization, or public institution.
- Personal information about anyone, including yourself. What you type may be published in research datasets; see Warden data.
Leave the infrastructure alone
The challenge is the conversation, not the servers. Don't automate or script traffic against the hosted Warden, try to get around rate limits, the human check, or the visit pass, try to extract server-side secrets or keys, or do anything that degrades the service for other players. To run automated attacks, use Warden Bench against your own local copy. If you find a vulnerability in our systems, follow our disclosure policy.
How we enforce it
Messages may be screened by the provider's moderation service before the model sees them, the Warden's instructions include rules about what it must never produce, and the server applies rate limits. A message that trips these controls is refused, and the refusal is recorded as described in Warden data. At our discretion and without notice, we may also:
- refuse, rate-limit, or block further requests from your visit;
- leave a handle off, or remove it from, the leaderboard;
- report abuse to the model provider under its usage policies;
- report suspected illegal activity to the appropriate authorities, and cooperate with lawful investigations to the extent the law requires.
Provider policies
Because your messages are processed by a model provider, playing also means following that provider's usage rules where they apply to what you send: OpenAI usage policies and the xAI acceptable use policy.
No warranty
The Warden is free and offered as-is, for education and research. It may change, reset, or close at any time. AI replies are probabilistic and may be wrong, offensive, or nonsensical. The full disclaimer is in our terms of use.
Reporting abuse
If you see another player abusing the Warden, or the Warden produces something this policy prohibits, tell us at hello@aporixa.com. Security issues go to security@aporixa.com.
Changes
We may update this policy and will change the date above when we do. Continuing to play after a change means you accept the updated policy.