Adversarial Chatbot
Talk to a model protected by instructions and filters and get it to reveal what it should not. In the browser, nothing to install.
AI security, hands-on
Learn AI security by breaking real models and agents. From guided prompt injection to compromising systems where the LLM is the access vector, in a safe and controlled environment.
Challenges
Hackers
Flags Sent
What the platform offers
Challenges on real systems
Every challenge runs against a working language model or agent, not a simulation. The techniques are practised the way they are applied in an audit.
Isolated environments
Each instance is deployed individually, with no Internet egress and a limited lifetime. It can be attacked without risk to third parties.
Measurable progression
Challenges by level, from introductory to advanced, with a ranking, achievements and writeups to check your own work against.
Two challenge formats
Machines
Full systems reachable over VPN where an AI agent is the way in: compromise it, gain execution and escalate privileges.
Registration is free and requires no card. The Pro plan extends access to the whole catalogue, with no lock-in.