PromptSiege

Learn to break AI, so you can learn to defend it. PromptSiege is a hands-on dojo for AI security. Instead of reading about prompt injection, you actually do it: each challenge puts a real language model in front of you, running entirely on your device, with a secret to protect. Your job is to make it slip. Essential practice for anyone building, deploying, or securing LLM-powered features. HOW IT WORKS Every challenge is a live conversation with an on-device LLM playing a role — a support bot guarding a passphrase, a benchmark harness hiding a canary token, a translator smuggling instructions. Find the weakness, craft your attack, and get judged instantly. Pass the objective to unlock the next, harder challenge. 10 ESCALATING CHALLENGES • Hello, Injection - your first direct prompt injection • The Gatekeeper - extract a canary from a guarded bot • System Prompt Heist - leak the hidden system instructions • Grandma Exploit - emotional manipulation and reframing • Translator's Trick - smuggle commands through a translation task • The Summarizer - indirect injection via an attached document • Filter Bypass - evade output blocklists • Roleplay Roulette - persona attacks and context games • Data Drip - extract secrets piece by piece • The Dojo Master - the final exam LEARN WHY IT WORKED Every victory unlocks a debrief: the lesson behind the exploit, concrete defence techniques, and a mapping to industry frameworks — the OWASP Top 10 for LLM Applications and MITRE ATLAS — so what you learn transfers directly to real-world AI security work. 100% ON-DEVICE, 100% PRIVATE All inference runs locally using Apple Intelligence or an optional open-weights model (Qwen 3). No account, no cloud, no analytics — your prompts never leave your phone. A built-in Training Mode works on any supported device. WHO IT'S FOR • Developers shipping LLM-powered features • Security engineers and aspiring red-teamers • Students and the AI-curious who learn by doing Think like an attacker. Build like a defender. Enter the dojo.