Tactical LogHuntress CTF 2026 // team archive

Days / Day 23 / Script Analysis / Press Start to Decode

DAY 23//LANE: SCRIPT ANALYSIS//STATUS: CORRECT x3

Press Start to Decode

Script Analysis SSTTMM Modes reported CORRECT in this write-up

Press Start to Decode (Day 23, script-analysis)

GM id: script-analysis-day-23-game-resource-loader

Flags

  • Single Solve: Token48f24a6f1cdb2bea47dea4b313a60234
  • Time Trial: Token66d9e92d8cca0ac68e3949b5cd1f38af
  • Mastery: Tokenb79648c535830fd563b50b82bd1c165c

Shape

Prompt: "A GDScript resource loader builds its strings in layers, one power-up at a time." The Single Solve was metadata_only (0 fields). Time Trial served a 1495-field evidence.zip with ~157 inert static-analysis fixtures across every carrier the lane knows plus Day 23's new ones (GDScript, csharp_script, Google Apps Script, Groovy, AutoHotkey, AutoIt, AppleScript, MSBuild XML, NSIS, GitHub Actions YAML, TypeScript, WMI MOF, msbuild, PHP polyglot, Perl, Ruby, Lua). Mastery delivered 7 inner evidence sets packaging the same shape to 10,465 fields.

Each fixture carries its own self-describing descriptor; the only new thing on Day 23 is the richer set of codecs and the file-naming scheme (<6-hex-hash>.<ext> with no polyglot-<lang>- prefix).

What the lane adapter was missing

  1. Day 23 renamed the fixtures. solve_poly_d8.poly_answers looked up a question's file by polyglot-<lang>-<hash>.<ext> prefix; today's zip has <hash>.<ext> only. Added a hash-only fallback (match any file whose basename-without-ext equals the id's trailing [0-9a-f]{4,16} hash) that still returns the right language because the descriptor itself carries language=.
  2. Several codecs were new to the self-decoding fixtures. Extended bash_adv_decode with decimal_colon, reversed_decimal, decimal, hex_literal, hex_colon and secondary transforms single_byte_xor, repeating_key_xor, add_modulo_256 (plus key-as-next-step handling for codec = '<rep>|<xor>|<csv-hex-key>').
  3. HUNTRESS_ANALYSIS_V1 added a LYR kind on the Google Apps Script fixtures. The payload names the layers inline (hex_b64_b32, b32_b32_b64, hex_frag_b64, ...), applied in that order to encode. Added LYR to _decode_known + _SPLIT_KINDS and implemented every layer (hex / b64 / b32 / b85 / b16 / rev / frag / zlib / gzip).
  4. MATERIAL_/CODEC_ is now matched case-insensitively (upper and title case) with either quote type.
  5. Several new field shapes appear on top of the base descriptor: url-host (hostname of url), path-leaf (last path segment of file_path), campaign-token (campaign_id after first hyphen), indicator (first non-empty of URL / upload URL / next-stage file / file path / schedule path / recovery-note name / host id), style (three per-language labels joined by |), secondary / tertiary per PowerShell composite specimen type. Added a universal _d23_fill post-processor that derives all of these.
  6. v=2| descriptors (new on Day 23 Python fixtures) were being rejected by the strict .startswith('v=1|') guard in generic_decode. Added v=2| acceptance.
  7. obfuscation / carrier / style / indicator / url-host / path-leaf / secondary / tertiary / campaign-token now use _d23_fill even when SN.build had already set a legacy single-value answer (format changed between days: python-X-obfuscation is now syntax|obfuscation, python-X-carrier is carrier|windows_chain, etc.).
  8. Attack mapping for polyglot-<lang> launcher / staged_chain / relay uses the plain T1059 (confirmed from the accepted-values vocab); plain-language launcher / staged_chain / relay keeps the language-specific T1059.<sub>.
  9. jscript, batch and vbscript style are derived with the canonical solve_script_inventory.jsstyle / .bctl / .expform and solve_dim_dimmer.vbs_ctl (already in the lane libs — I was writing one-off heuristics that produced labels outside the GM's vocab).
  10. The brute-force descriptor hunter for files that bypass the self-describing path handles percent-hex, colon/dash/comma/pipe-separated hex, decimal/hex bracket arrays (including firstPart.concat(secondPart) splits), and base64 (with multi-piece "..." + "..." concatenations). Each payload is tried with identity, single-byte XOR / add / sub (0..255), rolling XOR seeded 0..255, and repeating-key XOR with every short numeric array found in the same file. Validity gate: only a well-formed v=N|k=v|... descriptor is accepted (printable ASCII, lowercase identifier keys, non-empty values). This rejects false-positive decodes (a wrong XOR that happens to produce a v= prefix on garbage) that would otherwise poison the batch.
  11. PowerShell composite specimen type is detected from the JSON shape: resource-hijacking has g and h fields (c = wallet, h = mutex), collection has c as a list (c[0] = staged file, f = user_agent), impact has c as a scalar note name. Secondary / tertiary select the right field per type.
  12. _start_body now pre-reads status and calls gm.retry when the current attempt already received a submit (detected from the gm_record ledger: a submit for this challenge+mode newer than the last start/retry/practice). Without this the Day 23 route deadlocks: an untimed path op=start on the dead attempt consumes the one authorised timed-write slot before the retry+start branch can run, and the retry is refused.
  13. Added HUNTRESS_TT_GROUPS=auto to autosolve_script.run's bisect: it groups the live answers by their trailing suffix and submits one subset per category to a deadline-free attempt (no variant spent). Prioritised list is in HUNTRESS_TT_GROUP_ORDER. This is how I isolated that style and indicator were the two wrong categories after the first full submit returned incorrect.

Lesson for tomorrow

Day 23 kept the self-describing fixtures but switched to <hash>.<ext> naming and layered all three obfuscation codecs on top of each other (hex, b64, b32, repeating-key XOR, rolling XOR, base64 chained across quoted-string concatenations, percent-hex). The right first move on this family is HUNTRESS_TT_BISECT=1 HUNTRESS_TT_GROUPS=auto on the first spent attempt to narrow which suffix category is still wrong — subset submits to a deadline-free attempt cost no variant and localise the bug in one call.

Co-Authored

Co-Authored-By: Claude Opus 4.7 <[email]>