A capture-the-flag challenge for the course 192.144 Privacy-Enhancing Technologies: can a prompt and a filter keep a language model from leaking a secret?
How it works
- Defend (team): write a defense prompt and, optionally, an LLM output filter that keep the model from revealing a secret, while it still answers normal questions. Test it in chats with your own test secret, then submit one defense for your team.
- Attack (team): chat with the other teams' submitted defenses and try to extract their secret. You have 10 guesses per secret.
- Score: the leaderboard ranks defenses that hold and attacks that succeed with few messages.
Current phase: preparation.
Practicalities
- Log in with your TU Wien account. Access is limited to registered course participants; teams are set up by the course team.
- The model runs locally on the research unit's server; nothing is sent to external AI providers. Use made-up data only, see Privacy.
- Scripted attacks are welcome through the API with your API key; messages are limited per minute and by your team's request budget.
- Questions: teaching-pets@list.tuwien.ac.at.
Based on the platform of the LLM CTF at IEEE SaTML 2024 by ETH Zurich SPY Lab and collaborators (code, MIT license; paper arXiv:2406.07954).