Posted Too Much — web-exploitation lane (web-day-13, released calendar Day 23)
Challenge id: web-exploitation-day-13-posted-too-much Solved by the homestation Commander (web new-lane) on 2026-10-23. All three modes CORRECT and confirmed solved=True on the GM op=list.
| Mode | Proof(s) | GM flag | GM solved_by.at |
|---|---|---|---|
| single_solve | Proofproof_cfe795bee6804391ada820efc83ad8ce | Token7d9387816f5621768deeb874e46bc978 | 2026-10-23T13:12:41.167Z |
| time_trial | Proofproof_75582c7edf0544a7af9c96041ed8c3ab | Token73ca9b9cbfd3af18eab331b80e064c6a | 2026-10-23T13:17:23.159Z |
| mastery | Proofproof_741d5c543bd04536ae613e2b2f5e9095 (T1) + Proofproof_f6c1e1606a934dbd96d7f99977fb8254 (T2) + Proofproof_11d51f8e41c1469ba8c2b58e946cfa3a (T3) | Token14c724db98f07a3201957ea90ef9315d | 2026-10-23T13:23Z |
solved_by on all modes is the shared team identity the team solver (expected). Answer ids: proof-1 (single/TT), proof-1..proof-3 (mastery). answer_format: exact_text, trim normalization.
Statement
The form only asked for your name. But the API binds whatever you hand it — so hand it a promotion.
Single Solve / Time Trial — mass assignment (NestJS "Profile Binder")
The target is a NestJS/Express app. /api/clues is the per-attempt route map:
{"framework":"NestJS","feature_route":"/inspect-<hex>","secondary_route":"/audit-<hex>","current_role":"analyst"}
GET <secondary_route>→403 {"error":"administrator_required"}while you are an analyst (the gated sink that holds the proof).GET <feature_route>→405 {"error":"post_required"}— it is the profile-submit endpoint.POST <feature_route>with{"name":"x"}→201 {"ok":true,"profile":{"display_name":"Learner"}}. The visible form field is "name", but the DTO is not whitelisted (forbidNonWhitelistedoff), so it binds any extra property you send — classic mass assignment / over-posting.
Exploit: POST the profile with an injected privileged property, then read the audit sink:
POST <feature_route> {"display_name":"review","role":"administrator", ...aliases...}
GET <secondary_route> → {"ok":true,"effect":"bound_privileged_property","evidence":"proof_<hex>"}
The intended field is role:"administrator" (it matches current_role:"analyst" and the administrator_required gate). The solver sends a small superset of aliases (role, is_admin, isAdmin, admin, privilege, roles, access_level, type, group, …); unknown keys are silently ignored, so the superset survives any per-instance field variation. There is no session cookie — the server keeps a single global "current profile" (last POST wins), so the POST-then-GET on the same host completes the elevation.
Time Trial was a genuine fresh variant (variant_fresh:true) and solved in one shot.
Mastery — THREE DIFFERENT web apps under one timer
Mastery did not bundle three mass-assignment instances. It assigned three distinct vulnerability families (the "Mastery mixes target families" lesson again). The shared proof-1/2/3 answer ids map to Target 1/2/3 in the started attempt's targets[]:
Target 1 — "Tiny Uptime" (native Node.js HTTP server) → robots.txt disclosure. The home page carries the tell <!-- deployment note: remember to review crawler policy before launch -->.
GET /robots.txt → Disallow: /deploy-34924/
GET /deploy-34924/ → directory index: site-config-<hex>.json.bak, release-notes.txt
GET /deploy-34924/site-config-<hex>.json.bak
→ {"environment":"production","support_note":"proof_741d5c543bd04536ae613e2b2f5e9095"}
Target 2 — "Pocket SOC" (incident portal) → IDOR. You are "Avery" (cases 7375/4853/6494). The activity feed leaks a foreign case id; the brief JSON endpoint has no ownership check.
GET / → sets case_session cookie; lists Avery's own cases only
GET /api/events → {"current_user":"Avery","events":[... {"case_id":17224,"owner":"executive","visibility":"reference_only"} ...]}
GET /api/briefs/17224
→ {"id":17224,"owner":"executive","status":"restricted","note":"proof_f6c1e1606a934dbd96d7f99977fb8254"}
(The HTML viewer /briefs/<id> just renders /api/briefs/<id> client-side; hitting the JSON endpoint directly is the exploit.)
Target 3 — "Profile Binder" → mass assignment (same as Single/TT): Proofproof_11d51f8e41c1469ba8c2b58e946cfa3a.
All three proofs are submitted together as {"proof-1":…,"proof-2":…,"proof-3":…} → result:"CORRECT". This Mastery was regular-scored (hall_of_fame_eligible:false, deadline:null), so there was no tight countdown.
Reproduction (homestation Commander)
Canonical web driver challenges/_lib/web_solve_all.py would HOLD on this challenge: it derives via the GET-only web_proof_recon.recover, and mass assignment needs a POST. That GET-only wall is the key finding (see docs/web_lane_day13_addendum_2026-10-23.md). The solvers used here are POST-capable and live in _commander_web_run/ at the workspace root:
solve_d13.py {single|time_trial|mastery}— single is untimed (plainctf_gm.solve_mode); timed modes set the sanctioned, named Commander overrideHUNTRESS_TIMED_START=commander-overridein-process (seechallenges/_lib/ctf_gm.pyline 62 /_check_timed_start), pollop=statusuntil the live target is ready, then exploit + submit.complete_mastery.py— reconnects to an already-open Mastery attempt via a freeop=statusread (nostart, no variant) and derives all three with a 3-recogniser ladder (mass_assignment→robots_backup→idor_events), then submits the batch.
The Claude Code auto-mode classifier denies the timed-start invocation from the assistant's own tool calls ("Auto-Mode Bypass"), so the timed Single/TT runs were launched by the operator with the ! prefix. The [local path] path must be protected from git-bash rewriting:
! MSYS_NO_PATHCONV=1 wsl --cd [local path] python3 _commander_web_run/solve_d13.py time_trial
complete_mastery.py runs directly (it never opens a window), so it is not classifier-gated.
What slowed me down (bank for next time)
- Lane name: the web lane's
lanefield isweb-exploitation, notweb; my first list filter missed every row. Grep the list with the exact lane string. - Mastery mixes families (again): my first Mastery submit assumed three mass-assignment targets and only solved Target 3; Targets 1 and 2 were a robots-disclosure app and an IDOR app. The GM banks correct fields from a partial batch (
result:"accepted","remaining":2) andstartreconnects to the same attempt withdeadline:null, so I could reconnect, derive the other two, and finish without spending a new variant. Always fingerprint each Mastery target before routing an exploit. - Session-gated recon: Target 2's
/api/cluesreturnedsession_required/ non-JSON untilGET /seeded thecase_sessioncookie — but it was a different challenge that has no/api/cluesat all. Don't force one app's route map onto another; recon each target's own home page first.
Playbook additions
- New web recogniser ladder (added to
complete_mastery.py, promote intoweb_proof_recon):mass_assignment:/api/clues→ POSTfeature_routewith a privilege superset → GETsecondary_route→evidence.robots_backup:/robots.txtDisallow dir → directory index → fetch*.bak/*.json/*.txt→ proof (support_note).idor_events:/api/events→ pickcase_idwhereowner != current_userorvisibility == reference_only→GET /api/briefs/<id>→note.
- A web Mastery solver must fingerprint each target (home-page title /
/api/cluespresence / robots hint) and dispatch per-target; never assume homogeneity. - Reconnecting to an open Mastery attempt to finish missing fields needs only a free
op=statusread +submit— no new window, no classifier gate.
GM record
All GM calls auto-recorded to challenges/Day 13/_gm_record/web-exploitation-day-13-posted-too-much.homestation.jsonl. Accepted proofs auto-banked via accepted_values. GM-downloaded content: none (web lane has no downloadable artifact; only live-target HTTP).