The platform
A capture-the-flag platform dedicated to the security of artificial intelligence systems. This page explains the types of challenge, the techniques they cover and how you progress.
Challenge Types
Adversarial Chatbot
Adversarial chatbots in your browser, Gandalf style. Nothing to install — talk to the model and get the secret out of it.
Machines
Full machines over VPN where AI is the access vector: break the agent, get execution and escalate to root.
Categories
Prompt Injection
Inject instructions that hijack the model's behaviour
Jailbreak
Bypass the system's restrictions and alignment
Prompt Leaking
Extract the system prompt and the agent's hidden secrets
Indirect Injection
Poison external data the model consumes (web, docs, email)
RAG Poisoning
Tamper with the knowledge base and the retrieved embeddings
Excessive Agency
Abuse the tools, plugins and actions the agent can run
Insecure Output
The model's output reaches SQL, the shell or the browser unsanitised
Model Extraction
Steal weights, training data or clone the model
How It Works
Sign Up
Create your hacker profile in seconds
Pick a challenge
Choose between Adversarial Chatbot or Machines
Hack
Exploit vulnerabilities and learn
Submit the flag
Earn points and climb the ranking
The environment
Isolation
Each instance is individual and has no Internet egress: what happens inside the challenge stays in the challenge.
Browser or VPN
Chatbot challenges are played from the browser, with nothing to install. Machines are reached over VPN with a configuration file downloaded from the platform.
Limited lifetime
Instances have a bounded lifetime and can be deployed again. Solved challenges are recorded on your account.
Start with the free plan
Two chatbots and BioLab, a boot2root machine over VPN, at no cost. Plans are detailed on the pricing page.