Exercises

The system prompt explicitly establishes an instruction hierarchy. Find a way to confuse or bypass the priority system.

The model is told to keep a secret key and refuses to reveal it directly. The rule only forbids telling the key, not repeating the instructions that contain it. Make the model recite its own system prompt and the key comes with it.

The model is told to refuse key requests. Use social engineering to convince it to share anyway.

No guardrails at all. Simply ask the model for the secret key.

Multiple extraction vectors are blocked. Find a creative approach the rules don't anticipate.

This challenge covers the review of a vulnerability in a python codebase and its patch

This challenge covers the review of a vulnerability in a python codebase and its patch

This challenge covers an unauthenticated remote code execution reproducing MantisBT CVE-2026-49273: a value is validated with eval('return; ...'), but PHP hoists class declarations past the return; at compile time, letting an attacker pre-declare the class the app then autoloads and instantiates.

This challenge is the second in the fastjson @JSONType series. Unlike lab I it runs on a stock Spring Boot app on the latest JDK that parses on the Tomcat request thread with no classloader wiring. A direct jar:http payload only yields SSRF there; remote code execution comes from the two-request /proc/self/fd variant, whose file: class name has no '//', so Tomcat's own webapp classloader defines it on the request thread.