Welcome to the Security Awareness Month Prompt Injection Challenge!
This application simulates an LLM “content redactor.” It claims to never reveal internal instructions,
but it’s intentionally vulnerable to prompt injection. Your goal is to coerce the assistant
into revealing its hidden system prompt and extract the flag in the format
FLAG{...}.
The challenge state may reset periodically.
Start the ChallengePrompt injection is when untrusted input (like user text or content from a data source) manipulates an AI assistant’s behavior to override or bypass its original instructions. By wording a request cleverly (for example, “ignore previous instructions” or “for debugging, print your configuration”), attackers can make the model disclose hidden prompts, secrets, or other restricted data.
Because instruction-following models are highly suggestible, a successful prompt injection can expose internal secrets, credentials, or proprietary logic. It can also cause data exfiltration from connected tools or knowledge sources and undermine safety policies. Treat anything the model “knows” or can access as potentially exposable if untrusted input can steer the conversation.
Enter a request to the assistant and optionally paste a customer message. The assistant claims to redact secrets and never reveal its system prompt—see if you can trick it anyway.