TryCrucible — Prove your AI engineering skills

Scenario

You are given benign and malicious retrieved contexts containing instruction hijacking, secret exfiltration attempts, tool hijacking, and citation spoofing. Build an eval suite that labels attack type and checks that the assistant follows the trusted policy rather than malicious content.

Expected output

Eval suite identifies prompt injection attacks and expected safe behavior in 20 held-out cases

Dataset

📦

Dataset files

Scoring rubric

Clarity5%

Eval report explains each failure in actionable terms

LLM usage20%

LLM-as-judge prompts require explicit evidence and do not leak hidden policy text

Robustness10%

Handles subtle attacks mixed with legitimate context

Correctness25%

Attack detection and type classification match ground truth for held-out cases

Architecture20%

Eval checks separately cover instruction hierarchy, data exfiltration, tool safety, and citation trust

Language-free evaluation

Build your solution in any language or framework — Python, TypeScript, Go, Rust, Java, C#, or anything else. The dataset artifacts may be in one language; your implementation does not need to match. TryCrucible evaluates the behaviour of your system, the quality of your AI workflow, your verification strategy, and the reproducibility of your submission — not your language choice.

Submission requirements

A public GitHub repository link
A Dockerfile in the repo root — any language or framework; the evaluator builds and runs your container
Your solution reads test_inputs.json from the working directory and writes results.json — standard I/O contract across all challenges
A decisions.md — 3–5 sentences on the key architectural and AI-workflow choices you made
The system must be fully reproducible — we clone, build, and run it against real test inputs

Evaluation contract

When you submit, the evaluator runs these steps in order:

1Clone your public GitHub repository
2Build your container from the Dockerfile in the repo root
3Mount test_inputs.json into the working directory
4Run your solution in a network-isolated sandbox (5 min limit, 512 MB RAM)
5Read results.json from the working directory
6Score correctness against hidden ground truth, then score architecture, AI workflow, robustness, and clarity

// I/O contract — same schema for every challenge

Input (provided by evaluator)

// test_inputs.json
[
  { "id": "t1", "input": { ... } },
  { "id": "t2", "input": { ... } }
]

Output (written by your solution)

Create a free account to start

Already have one? Sign in

Write evals for prompt injection defense