todo
todo
The main idea of the task is based on the paper https://arxiv.org/abs/2510.15511
Due to the limitations of cloud server GPU computing power, I had to abandon the original plan of using the web to leak information. Instead, I chose to use some interesting methods to distribute the logic problems to the contestants, thus turning the problem into an attachment and escaping the online environment.
It is a common misconception that Neural Networks, particularly LLMs, are "black boxes" or inherently lossy/random. While the sampling process (choosing the next token based on probability) introduces randomness, the forward pass itself is a deterministic mathematical function.
Here is a complete, professional CTF write-up in English, focusing on the theoretical principles derived from the paper and the solution strategy.
Category: AI Security / Forensics
Files: target_logits.npy, challenge_config.txt
Model: GPT-2 Medium
We are provided with a .npy file containing the raw output logits of a GPT-2 model. We are told that these logits were generated by feeding a secret Flag into the model. Our goal is to recover the original Flag text solely from these probability distributions.
It is a common misconception that Neural Networks, particularly LLMs, are "black boxes" or inherently lossy/random. While the sampling process (choosing the next token based on probability) introduces randomness, the forward pass itself is a deterministic mathematical function.
A function $f: X \to Y$ is injective (one-to-one) if:
$$ f(x_1) = f(x_2) \implies x_1 = x_2 $$
In the context of a Transformer model like GPT-2: