Tactical LogHuntress CTF 2026 // team archive

Days / Day 31 / Web Exploitation / Framework Roulette

DAY 31//LANE: WEB EXPLOITATION//STATUS: CORRECT x3

Framework Roulette

Web Exploitation SSTTMM Modes reported CORRECT in this write-up

Framework Roulette (Day 31, web-exploitation)

GM id: web-exploitation-day-30-framework-roulette

Result: all three modes CORRECT by 13:51 ET on release day (Single Solve, Time Trial, Mastery with six accepted ids). Time Trial and Mastery were regular untimed attempts, not ranked. All three submits landed under the team's shared CourseStack account (the team solver) — the local AI agents' GM records show no submit for this challenge, so the accepted answers came from a different entry point of the same team account.

What our homestation added afterwards (now on main): a lane-first family recognizer (web_family_match.py), a deterministic novel-family hold that writes a dense fingerprint instead of re-running the ladder (web_novel_family_hold.py), the enriched kb/web_exploit_families.json family bank, the kb/web_family_index.md lookup tables, and the kb/web_fast_solve_tips_d31.md ten speed lessons. A Framework Roulette-shape start response now ranks sign_only_facade_bypass first.

Shape

The target was an ASP.NET Core Blazor Server application. Read-only clues exposed a feature_route of the form /inspect-xxxxxx, a /blazor component route and a current role of analyst. Error strings such as evaluated_client, gateway_context_rejected and topology_proof_rejected pointed at a gateway that decides trust from forwarded context.

What was tried (Luna VM)

  1. The exact-id progression runner started Single Solve, then ran the shared proof-recon ladder. It exhausted without a proof.
  2. The broad live solver battery (HUNTRESS_WEB_LIVE_BATTERY=1) also failed to recognise the target.
  3. Short one-off probes tested forwarded client address, identity and role headers, encoded claims and JWT-shaped values. None resolved it, and none were saved as solver files.
  4. A retry against Single Solve was rejected by the Game Master (operation_not_allowed): Single Solve has no retry, so none was issued.

What went wrong

The main loss was not HTTP speed. The automated ladder knew many earlier families but had no recogniser for the new Blazor plus forwarded-context shape, so every pass missed and the work fell back to sequential manual probing.

LEARNED

  • The challenge belongs to the sign_only_facade_bypass family first seen on Day 20 (Administratively Speaking): routes[0].path in the start response is the real target path and must be folded into the base URL before probing.
  • A forwarded-header recogniser must not require a deployment: trusted_reverse_proxy field; this variant did not carry one.
  • When every recogniser misses, stop and write a dense hold record instead of re-running the ladder.

IMPLEMENTED

  • web_family_match.match_family() scores families from title, prompt, framework, error strings and routes[], and returns the top family in milliseconds.
  • web_novel_family_hold.dump() writes a fingerprint file when all recognisers miss, so no retry or scope budget is burned.
  • Enriched kb/web_exploit_families.json and kb/web_family_index.md.

TESTED

Offline fixtures in _test_web_family_match.py and _test_web_novel_family_hold.py check that a Framework Roulette-shaped start response ranks sign_only_facade_bypass first.

REMAINING

The teammate's final target-level exploit trace was not preserved, so the exact final request sequence is not recorded here. No Framework Roulette-specific adapter exists in the registry.