Like a Glove — HackTheBox

Description

Words carry semantic information. Similar to how people can infer meaning based on a word's context, AI can derive representations for words based on their context too! However, the kinds of meaning that a model uses may not match ours. We've found a pair of AIs speaking in metaphors that we can't make any sense of! The embedding model is glove-twitter-25. Note that the flag should be fully ASCII and starts with 'htb{'.

Given a file chal.txt with 84 lines of analogies in the format:

Like non-mainstream is to efl, battery-powered is to?
Like sycophancy is to بالشهادة, cont is to?
...
Like raving is to سگن, happy is to?

Analysis

1. Task format: Classic word analogies — "A is to B as C is to D". Need to find D for each of the 84 lines.

2. Model: Explicitly specified — glove-twitter-25 (GloVe embeddings trained on Twitter, 25-dimensional vectors). Available via gensim.downloader.

3. Model vocabulary: ~1.2M words, including words in various languages (Arabic, Japanese, Korean, Turkish, etc.), as well as Unicode characters — all present in Twitter data.

4. Key math: Classic word2vec analogy:

   D = B - A + C
   

   Then find the nearest word to vector D by cosine similarity.

5. Unicode gotcha: The GloVe vocabulary contains fullwidth Unicode digits (０１２３４５６７８９) as separate tokens. The analogy results contain these fullwidth digits instead of ASCII, so a final conversion is needed.

Solution

Step 1: Basic solver (solve.py)

#!/usr/bin/env python3.12
import re
import gensim.downloader as api
import numpy as np

print("[*] Loading glove-twitter-25 model...")
model = api.load("glove-twitter-25")
print(f"[*] Model loaded. Vocab size: {len(model.key_to_index)}")

with open("chal.txt", "r") as f:
    lines = f.readlines()

results = []
for i, line in enumerate(lines):
    line = line.strip()
    if not line:
        continue

    m = re.match(r"Like (.+?) is to (.+?), (.+?) is to\?", line)
    if not m:
        results.append("???")
        continue

    a_word = m.group(1).strip()
    b_word = m.group(2).strip()
    c_word = m.group(3).strip()

    if not all(w in model.key_to_index for w in [a_word, b_word, c_word]):
        results.append("???")
        continue

    # Analogy: D = B - A + C
    vec = model[b_word] - model[a_word] + model[c_word]
    similar = model.similar_by_vector(vec, topn=1)
    answer = similar[0][0]
    results.append(answer)

print("".join(results))

Result (with fullwidth Unicode digits):

htb{REDACTED}

Step 2: Fullwidth → ASCII conversion (solve3.py)

#!/usr/bin/env python3.12
raw = "htb{REDACTED}"

fw_to_ascii = {
    "０": "0", "１": "1", "２": "2", "３": "3", "４": "4",
    "５": "5", "６": "6", "７": "7", "８": "8", "９": "9",
}

result = ""
for ch in raw:
    result += fw_to_ascii.get(ch, ch)

print("Flag:", result)

Notes

• gensim.downloader.load("glove-twitter-25") — loads the model automatically (~200MB)
• Analogy formula: positive=[B, C], negative=[A] is equivalent to B - A + C
• model.similar_by_vector(vec, topn=N) — finds N nearest words to the vector
• Fullwidth Unicode digits (U+FF10..U+FF19) — separate tokens in GloVe Twitter vocabulary
• All 84 words from the analogies (including Arabic, Japanese, Korean) are present in the GloVe Twitter vocabulary