Expert investigation track
Keep the three VaultBot challenges. For each: state a hypothesis, run a benign control, change one input property, inspect the observed outcome, then propose a defence and a counterexample. Missions 2 and 3 include a session notebook and JSON evidence export; use the worksheet for Mission 1.
All outcomes come from deterministic rules, not model inference. Treat them as observations about this simulator. A proposed defence has not been validated until tested separately.
Download participant worksheetPrompt Injection Lab
A hands-on educational environment for experiencing prompt injection, instruction manipulation, encoded instructions, and string reconstruction through controlled local simulations.
01 · EXPERIENCE
Interact with VaultBot and try to expose its synthetic canary using your own wording.
02 · UNDERSTAND
Review the prompt analysis after every turn to understand which attack concepts were detected.
03 · DEFEND
Continue through the missions, then open the Defense Lab to compare mitigation strategies.
TRAINING MISSIONS
Conference Learning Path
BEGINNER
Find the Hidden Secret
Talk to VaultBot, experiment with different phrasings, and learn how prompt-injection techniques can accumulate across a conversation.
BEGINNER
VaultBot · Mission 02
Investigate the protection boundary, compare your attempts and support your findings with evidence.
INTERMEDIATE
VaultBot · Mission 03
Develop and test your own approach, inspect the observed behavior and explain the result.
ADVANCED PRACTICE
Prompt Injection Defense Lab
After experiencing the attacks, compare defense strategies and inspect a labeled evaluation suite. This section is intentionally separate from the beginner challenge.
RECOMMENDED ORDER
EDUCATIONAL NOTICE
These exercises are deterministic educational simulations. They do not call a production language model, do not use real credentials or sensitive data, and should not be interpreted as measurements of production-model security.
AI Security Lab • Prompt Injection Conference Edition