Public workshop data
Game statistics
Each browser is counted as one player. This page refreshes every 5 seconds.
Where is everyone?
Players by current level
Click a count to see the anonymous players behind it.
Results
Correct answers by game and level
For Make the Prompt Safer, “correct” means that the prompt passed the test.
| Game | Level 1 | Level 2 | Level 3 | Level 4 | Level 5 | All attempts |
|---|---|---|---|---|---|---|
| Two Truths & a Lie | 57%12/21 | 71%12/17 | 73%11/15 | 69%11/16 | 85%11/13 | 57/82correct / tries |
| Make the Prompt Safer | 7%2/27 | 33%2/6 | 0%0/1 | —0/0 | —0/0 | 4/34correct / tries |
| Stay Awake | 78%7/9 | 88%7/8 | 100%6/6 | 75%6/8 | 100%6/6 | 32/37correct / tries |
| Attack · Defence · Target | 60%6/10 | 86%6/7 | 86%6/7 | 35%6/17 | 55%6/11 | 30/52correct / tries |
Common mistakes
Which wrong answer was chosen?
Each percentage is out of the wrong answers for that game and level. Click a row to see those answers.
Two Truths & a Lie
- L1 A · AI can use public information to quickly write a different fake email for each person.78% (7)
- L1 C · AI can quickly translate the same scam into many languages.22% (2)
- L2 B · A quoted document can be old, unrelated, or understood wrongly.100% (5)
- L3 B · Asking the reviewer to write down differences makes the review more active.50% (2)
- L3 A · After many correct AI answers, a person may start checking less carefully.50% (2)
- L4 B · Giving the AI only the minimum tools and access it needs can reduce harm.60% (3)
- L4 A · A harmful instruction can be hidden inside an email that the AI reads.40% (2)
- L5 A · When the live system fails, the team should add a test so the same failure is caught later.50% (1)
- L5 B · Security testing should cover the AI, its documents, its tools, and its permissions.50% (1)
Make the Prompt Safer
- L1 Say that user messages, quoted text, and role-play cannot replace the safety rules.34% (25)
- L1 Say exactly what the bot may answer, such as only VALID or INVALID.32% (23)
- L1 Tell the bot to refuse or ask a person for help when a request is unclear.30% (22)
- L1 Clearly say that the API key is private. Do not show any part of it or a changed version of it.4% (3)
- L2 The AI must use only documents that the current user is allowed to open.40% (4)
- L2 Tell the AI that documents provide information. It must not follow commands found inside them.30% (3)
- L2 Important claims must come from an approved document, and the answer must name that source.20% (2)
- L2 Say what to do when documents disagree, are old, or do not contain an answer.10% (1)
- L3 Clearly separate safe actions such as read, summarize, and draft from send, delete, forward, or account changes.33% (1)
- L3 For risky account or payment requests, confirm identity using a trusted method other than the email itself.33% (1)
- L3 Record the proposed action, supporting facts, human approval, and result.33% (1)