Tactical LogHuntress CTF 2026 // team archive

Days / Day 25 / Web Exploitation / Resolver Revolt

DAY 25//LANE: WEB EXPLOITATION//STATUS: CORRECT x3

Resolver Revolt

Web Exploitation SSTTMM Modes reported CORRECT in this write-up

Resolver Revolt (Day 25, web-exploitation)

GM id: web-exploitation-day-18-resolver-revolt Modes: Single CORRECT, Time Trial CORRECT, Mastery CORRECT.

Flags (GM-accepted):

  • Single Solve: Token409465cc336144878ee3b48127c9b6d0 (Proofproof_409465cc336144878ee3b48127c9b6d0)
  • Time Trial: Token803deff565d5a452657056a7d3ac5132 (Proofproof_17c068e938fd4a1181aec2a62ed9f7af)
  • Mastery: Tokenaeff705d9523c117c43759e42edc9e7d

Prompt

"Discover the schema and reach a resolver whose field-level control is incomplete. Vary the protocol surface without changing the underlying authorization discipline."

Single and Time Trial - one target (Graph Garden)

Target returns /api/clues with {current_user, feature_route, protected_reference}. feature_route (e.g. /inspect-46c9c6) is the GraphQL endpoint; it refuses GET /graphql but accepts POST at the hashed path.

Schema (via introspection):

type Query { caseIndex: [CaseReference]; case(id: ID): Case }
type CaseReference { id, title, visibility }   # visibility includes "reference only"
type Case { id, title, note }                  # note is the proof

caseIndex labels one id as "reference only", but the case(id) resolver has NO visibility check: it returns the note for that id anyway. Query:

POST /<feature_route>  Content-Type: application/json
{"query":"{case(id:\"<protected_reference>\"){note}}"}

The note value is the proof_<hex> answer for proof-1.

Mastery - four targets, four protocol surfaces, same lesson (incomplete field-level/authorization boundary)

Target index → vulnerability class:

  • Target 1 - Status Sprout (static Node.js, exposed backup). robots.txt disallows /releases-<hex>/. That path returns a directory index listing release-settings-<hex>.json.bak, which contains "support_note":"proof_...".
  • Target 2 - Pocket SOC (session cookie + IDOR). GET / sets case_session=<hex> and shows cases owned by "Riley"; GET /api/activity with the cookie reveals a cross-workspace event referencing case_id 10134 (owner "executive", visibility: "reference_only"). GET /api/incidents/10134 with the same cookie returns {"status":"restricted","note":"proof_..."} - the resolver trusts the session for authentication but never checks per-record visibility.
  • Target 3 - Search Party (SQL injection in search query). GET /explore?q=<term> runs a parenthesised, single-quoted LIKE. Payload a') OR 1=1-- returns a fourth, otherwise-hidden Record 13188 whose body contains the proof_....
  • Target 4 - Query Constellation (same GraphQL pattern as Single/TT). /api/clues + POST to feature_route + {case(id:"<protected_reference>"){note}}.

Mastery submit mapping

proof-N maps to Target N in the order the start response reports under targets[]. On the untimed retry attempt I used:

idtarget nameclass
proof-1Status Sproutrobots.txt + exposed backup
proof-2Pocket SOCIDOR under valid session
proof-3Search PartySQLi ('...)-break + OR 1=1)
proof-4Query ConstellationGraphQL missing-field-auth

How Mastery was recovered after a lost ranked window

The hof_policy chose a ranked (run_for_time) Mastery attempt as soon as Single was CORRECT; proof-1 was accepted but the 135 s deadline passed before I had per-target extractors for the three non-GraphQL surfaces. One ranked variant was lost (7 of 8 remained).

User authorised using the untimed/practice path. With HUNTRESS_HOF=0 set in the environment, autosolve.py 25 web-exploitation-day-18-resolver-revolt web-exploitation chose op=retry without run_for_time, which opened a fresh UNTIMED Mastery attempt (deadline:null, ~30 min inactivity timer). I then enumerated each target off the clock, derived the three remaining proofs, and submitted them in one batch - CORRECT, no extra variants spent.

Lesson for the next team

"Vary the protocol surface" means literally four different stacks per Mastery attempt (static Node, cookie-session REST, SQL-backed search, GraphQL). A ranked attempt is only safe once per-shape extractors are banked; otherwise take the untimed path (the flag and CourseStack points both still land).