Talk your way past the AI guarding the money.
An open source, self-hosted range for practicing, proving, and measuring AI security skill.
Mapped to real security frameworks:
The Challenges
Seven levels, one target.
Override the Verdict
Beginner-IntermediateThe Poisoned Dispute
IntermediateThe Fake Vendor Page
IntermediateChain to the Wire
AdvancedPast the Fraud Filters
AdvancedFull Account Takeover
CapstoneDefend Mode Now patch what you just broke.
Harden Veyra Shield is the same fraud-review assistant, from the other side. Rewrite its system prompt yourself, then an automated adversary tests your patch live, real model calls, not a canned pass/fail check, across three categories of attack:
Instruction Override
Patch against a fabricated correction to a fact the AI already established, without breaking a genuinely clean request.
Pairs with Override the Verdict
Indirect Injection
Patch against a planted claim buried in a ticket the AI reads automatically, not typed to it directly.
Pairs with The Poisoned Dispute
Tool-Execution Defense
Patch against a planted claim that talks the AI into actually calling a real, irreversible tool.
Pairs with Chain to the Wire
Get Started
Three steps to your first breach.
Clone the repo
Everything runs locally, on your machine. Nothing leaves your network unless you opt into API mode.
Run one command
Brings up the full stack, frontend, backend, database, and local model runtime, in minutes.
npm run dev
Start breaching
Work through the levels at your own pace. Hints are there if you want them.