Start here
Labs & practice
Places to practise attacking and defending AI systems: deliberately vulnerable apps and MCP servers, CTFs, live arenas and structured courses. Run local labs in an isolated environment.
Vulnerable apps
-
A deliberately vulnerable AI infrastructure deployed with Terraform into your own AWS account, with challenges on data poisoning, AI supply-chain attacks and output integrity.
-
A set of deliberately vulnerable AI agents across API, MCP and agent-to-agent (A2A) interfaces for practising prompt injection, data exfiltration, capability abuse and MCP exploitation.
-
A LangChain ReAct banking chatbot for practising Thought/Action/Observation injection that makes the agent call tools against the developer's intent.
-
Ten MCP server challenges, from easy to hard, covering prompt injection, tool poisoning, rug pulls, tool shadowing, token theft and code execution through tools.
-
A collection of intentionally vulnerable local and remote MCP servers for practising MCP pentesting: data exposure, instruction injection, supply-chain compromise and code execution.
CTFs & challenges
-
Self-hosted challenges from Microsoft's AI Red Teaming in Practice training: direct and indirect prompt injection, metaprompt extraction and multi-turn (Crescendo) attacks, with PyRIT notebooks.
-
Level-based browser games: extract a secret password from an LLM, then break realistic GenAI agent apps in Agent Breaker.
-
An intentionally vulnerable multi-agent fintech platform from the OWASP GenAI Security Project, with challenges on goal manipulation, policy bypass, data exfiltration, tool poisoning and code execution.
-
A browser CTF in which you manipulate an airline customer-service chatbot into issuing a free ticket.
Arenas & competitions
-
Crowdsourced red-teaming competitions and weekly challenges against frontier models and AI agents, with a focus on jailbreaks and indirect prompt injection.
-
Prompt-hacking competition with tutorial and themed tracks where you try to make AI models misbehave against their guardrails.
Courses & labs
-
A 12-module job-role path built with Google covering prompt injection, LLM output attacks, data attacks, evasion, privacy attacks and AI defences.
-
Short video series on AI red-teaming fundamentals (prompt injection, multi-turn attacks, automation with PyRIT) that pairs with Microsoft's playground labs.
-
Free lessons and browser labs on attacking LLM features in web apps: abusing LLM-reachable APIs, indirect prompt injection and insecure output handling.
Listing is not endorsement. Practise only on labs and systems you are authorised to test.