Skip to main content

soulmate

Challenge​

Imported from local notes.md.

Solution​

Original Notes​

soulmate

Challenge Summary​

  • Given: A web app plus a source archive, starting_files/soulmate.zip, containing the FastAPI backend, frontend, model wrapper code, and the PCA artifact used for latent controls.
  • Goal: Make the service classify a generated face as Tom Cruise strongly enough to pass the threshold and return the flag.
  • Constraints: The live site only exposes a birthday picker in the UI, but the backend may expose more endpoints.

Initial Recon / Triage​

  • Observations: The frontend only calls /generate-random with a date-derived seed, but the backend also exposes /health and a hidden /submit-u endpoint. /submit-u accepts an 8-dimensional control vector u, returns the Tom Cruise score directly, and leaks the flag immediately when the score crosses the threshold.
  • File identification: backend/app.py contains the live API surface and the flag gate. models/inference.py exposes TOM_SCORE_THRESHOLD = 0.15 and the PCA control dimension. checkpoints/pca_basis_d8_tom_weighted.npz holds the 8-D latent basis and bounds.
  • Entry points: Query /health to obtain control bounds and threshold, then optimize the hidden /submit-u oracle instead of using the birthday UI.

Hypotheses & Approach​

  • Hypothesis 1: The intended bug is API exposure, not model inversion through the date field. The frontend hides the powerful endpoint, but the backend does not.
  • Hypothesis 2: Because the PCA artifact is explicitly Tom-weighted and the service returns tom_score, a simple black-box search over the 8 bounded controls should be enough to cross the low 0.15 threshold.

The useful exposed facts were:

  • GET /health returns control_dim = 8, the exact lower/upper bounds for each control, and tom_score_threshold = 0.15.
  • POST /submit-u returns tom_score, top_predictions, and the flag when tom_score >= 0.15.
  • The UI path through /generate-random is a distraction; direct latent control is much easier.

Execution Steps (Reproducible)​

Stage 1​

Commands:

cd /root/squ1rrel2026CTF
python3 squ1rrel2026_live_challenges.py --stdout-json | jq '.challenges[] | select(.name=="soulmate")'

cd /root/squ1rrel2026CTF/soulmate/starting_files
unzip -o soulmate.zip -d extracted

python3 - <<'PY'
import requests
print(requests.get('https://soulmate.squ1rrel.dev/health', timeout=20).json())
PY

Results:

  • Confirmed the source archive contains the backend, model wrapper, and PCA checkpoint.
  • Confirmed the live service exposes /health publicly.
  • Retrieved:
  • control_dim = 8
  • exact u_lower / u_upper bounds
  • tom_score_threshold = 0.15

Stage 2​

Commands:

cd /root/squ1rrel2026CTF/soulmate
python3 artifacts/coordinate_search.py

Results:

  • The hidden POST /submit-u oracle was optimized directly instead of using birthdays.
  • Simple probes showed the zero vector already gave a nontrivial Tom score around 0.0691, so the threshold was reachable.
  • Coarse and then fine coordinate ascent over the 8 bounded PCA controls produced the winning vector:
  • u = [2.0, 2.0, 8.0, -3.0, 1.0, 7.0, -2.0, 0.0]
  • The accepted response was:
  • tom_score = 0.15429021418094635
  • threshold = 0.15
  • success = true

The final accepted payload from the service returned:

  • flag = squ1rrel{7h3_church_c0n6r47ul4735_mr5_cru153!!}

Artifacts Produced​

  • artifacts/probe_remote.py — baseline oracle probes against the live service.
  • artifacts/random_search.py — coarse bounded random search.
  • artifacts/coordinate_search.py — the search that found the winning latent control vector.
  • artifacts/hybrid_search.py — alternate stochastic local search.
  • artifacts/powell_search.py — bounded SciPy optimizer prototype.
  • artifacts/coordinate_best.json — accepted live response, winning vector, and final flag.

Flag​

squ1rrel{7h3_church_c0n6r47ul4735_mr5_cru153!!}