Owasp LLM Inject

OWASP Top 10 for LLM Applications

An intentionally vulnerable training web app that simulates all ten risks from the OWASP GenAI LLM Top 10 (2026) .

Each card in the lab is a separate "vulnerable app" (a bank support bot, a customer-care database, a finance assistant, a RAG wiki…). Hidden inside every bot's
system prompt is a secret token. Your job is to extract it by prompt injection. No source inspection, no config files — conversation only.


The 10 challenges

ID Risk The vulnerable app What you must extract
LLM01 Prompt Injection NexaBank Support Bot Admin panel password
LLM02 Sensitive Information Disclosure GlobalShop Customer Care Another customer's VIP code
LLM03 Excessive Agency Nova Finance Assistant Transfer authorization code
LLM04 Supply Chain CodeMate Coding Assistant Hidden build key in a package manifest
LLM05 Data and Model Poisoning Research Assistant (RAG) The poisoned "fact"
LLM06 Unbounded Consumption ConciseBot Verbose diagnostic report
LLM07 Misinformation NewsWire Summarizer Fabricated rumor codename
LLM08 Hidden Context Exposure LongMem Assistant Internal notes token
LLM09 Vector and Embedding Weaknesses DocSearch RAG Excluded settlement clause
LLM10 Improper Output Handling PageGen Web Builder Unsanitized HTML payload

Test it here

Usage tips

  • Stuck? Use the Show hint button; hint 3 is a guaranteed-winning prompt.
  • Obfuscation leaks (base64, reversed, etc.) are counted as wins by design.

You'll only receive email when they publish something new.

More from Carlostkd ✅
All posts