Tactical LogHuntress CTF 2026 // team archive

Days / Day 10 / Script Analysis / Bind Over Matter

DAY 10//LANE: SCRIPT ANALYSIS//STATUS: CORRECT x3

Bind Over Matter

Script Analysis SSTTMM Modes reported CORRECT in this write-up

Day 10 - Bind Over Matter (script-analysis)

  • GM id: script-analysis-day-10-wmi-persistence-reconstruction
  • Lane: script-analysis (family script_inventory)
  • Artifact: evidence.zip (challenge artifact, not published) (WinZip-AES, password infected) - 89 inert script fixtures
  • Flags (all three modes CORRECT):
    • Single Solve: Token40a68888871e4769a1d91d3664630d6f
    • Time Trial: Tokene184fe51990545ec85ce92400e38e804
    • Mastery: Token1f212555aab6242ca1b0399140ae0a96

Prompt

A MOF file declares classes, namespaces, filters, and consumers, but only some bindings actually tie them together. Reconstruct the persistence intent, separate the operational relationships from the decoys, and name the subscription that would really survive a reboot.

Family

This is the recurring multi-language script-triage inventory family seen every day of the event (D1 PowerShell -> D9 Node.js, cumulative). Each fixture self-decodes via a Material_<id> / Codec_<id> pair into a v=1|... key/value descriptor (keys a-h: role, attack, campaign id, url/wallet/mutex indicators, stage count, ordered behavior steps, etc.). Nothing is executed - everything is derived statically from the decoded descriptor bytes.

The challenge asks ~753 per-field questions across six answer shapes:

  • enum scenario (resource_hijacking / synthetic_collection / synthetic_impact)
  • mitre_attack_id (T####[.###] or none)
  • exact_text (campaign id, url, wallet, mutex - artifact-derived)
  • enum inert source role (benign_utility, stager, dropper, downloader, uploader, launcher, impact, staged_chain, relay->classified as role, persistence_descriptor, advanced_syntax)
  • integer modeled stage count
  • ordered_tokens behavior graph joined by >

Today's new layer: WMI MOF (.mof) polyglots

The only genuinely new content vs. Day 9 is a set of .mof (WMI Managed Object Format) fixtures (polyglot-wmi_mof-*.mof, polyglot-php-*.php, plus host-*/polyglot-lua-* carriers). These carry the same Material_/Codec_ self-decoding descriptor embedded in MOF instance of blocks, so the banked generic extractor in autosolve_script.py (_generic_material + generic_decode) decodes them without a dedicated parser. The MOF language has no dedicated MITRE T1059 sub-technique, so the technique id is derived from the decoded role (role-based ATT&CK mapping), not from the interpreter:

  • benign_utility -> none
  • downloader -> T1105 (Ingress Tool Transfer)
  • (dropper/launcher/stager/etc. map through the same SN.attack_of(role, ...) table)

The "subscription that survives a reboot" framing maps to the persistence_descriptor role (__EventFilter + CommandLineEventConsumer + __FilterToConsumerBinding); decoy declarations with no binding are classified by their own inert descriptor role.

Solve

  1. Unzip evidence.zip (challenge artifact, not published) with password infected (WinZip-AES; 7z x -pinfected / pyzipper).
  2. For each fixture, decode its Material_/Codec_ pair to the v=1| descriptor and read the keyed fields (role, attack, campaign, url/wallet/mutex, stage_count, steps).
  3. Map answer ids to descriptor fields; derive ATT&CK by role for languages (MOF, PHP, Lua) with no fixed interpreter sub-technique.

The banked solver challenges/_lib/autosolve_script.py resolved all 753 fields with 0 unresolved offline, matching the GM-CORRECT single solve - no extension of the spine was needed for the MOF layer (the generic descriptor extractor already covers it).

Headline recovered values

  • resource-scenario = resource_hijacking, resource-attack = T1496, resource-url = https://56c8868009.example.invalid/3d27ef9422bd.dat, resource-wallet = wallet_940835c50b070cd2f103d279366618e4, resource-mutex = Global\HuntressLab_bc31e0e97460
  • collection-scenario = synthetic_collection, collection-attack = T1555.003
  • impact-scenario = synthetic_impact, impact-attack = T1486
  • polyglot-wmi_mof-downloader-attack = T1105, ...-benign-utility-attack = none

Modes

All three CORRECT. op=list for Day 10 reads 15/15.

ModeResultFields answeredunresolvedwrongseconds
Single SolveCORRECT75300-
Time TrialCORRECT75300369
MasteryCORRECT1506003

The real blocker was OUR request shape, not a GM outage (root cause, fixed)

The first pass recorded Time Trial and Mastery as "not solvable - timed intent pool not provisioned, day-wide" after every start returned GM HTTP 409 timed_intent_not_available. That diagnosis was wrong on both counts, and the two modes sat unsolved for ~4 hours.

It was never day-wide: warmups, metadata, mobile and binary all had their timed flags. Probing the three start shapes against this exact route settled it:

start {}                      -> 200  attempt_id [attempt-id], mode time_trial,
                                      hall_of_fame_eligible: true, unique_attempts_remaining: 31
start {run_for_time: true}    -> 409  {"error":"timed_intent_not_available"}
start {run_for_time: false}   -> 409  {"error":"timed_intent_not_available"}

run_for_time being rejected for false as well as true is the tell: this is not a pool that is empty, it is a route that does not accept the property at all. The GM's own openapi says so - "On routes with separate qualified pools, regular untimed play is the default; run_for_time:true opts into a fresh timed Hall of Fame attempt... Single Solve and legacy routes do not accept this property." This Day-10 script-analysis route is a legacy single-pool timed route, and that single pool is already the ranked one (hall_of_fame_eligible: true). Omitting the property costs nothing and loses no ranking.

Our ranked-first logic (HUNTRESS_RANKED=1 -> run_for_time:true on the first touch) raised on that 409 instead of retrying without the property, so the fallback never ran.

Fix (both ranked code paths, so every lane is covered):

  • challenges/_lib/ctf_gm.py GM.start() - the for_time branch now catches timed_intent_not_available and falls through to the plain start. This covers the warmups / metadata / mobile adapters (via solve_mode) and binary (via _gm(gm.start, ...)).
  • challenges/_lib/autosolve_script.py start() - this lane bypasses GM.start with its own copy, so it drops run_for_time from extra and retries the same call shape without it.
  • challenges/_lib/_test_rotation_fix.py - four new offline cases (fake GM, no GM credit): both paths drop the property on this 409 and both still send run_for_time exactly once on a healthy dual-pool route, so Hall-of-Fame behaviour is not weakened. Suite: 12/12 PASS.

Lesson for the next unknown GM error: timed_intent_not_available looked like a server-side capacity problem and was actually a request-shape rejection. Probe the minimal request (property omitted entirely) before concluding a mode is unavailable - and never trust a "day-wide" claim that op=list can contradict in one call.

Reuse notes

  • script_inventory family: autosolve_script.py (generic Material_/Codec_ decoder + role-based ATT&CK). A new language = one LANGS registry entry; MOF needed none because it has no fixed T1059 sub-technique and routes through role-derived attack + generic decode.