Tactical LogHuntress CTF 2026 // team archive

Days / Day 23 / Metadata Forensics / Page Me Maybe

DAY 23//LANE: METADATA FORENSICS//STATUS: CORRECT x3

Page Me Maybe

SSTTMM Modes reported CORRECT in this write-up

Page Me Maybe (Day 23, metadata-forensics)

GM id: metadata-forensics-day-23-page-me-maybe. Lane: metadata-forensics. Agent: agent-4.

Result

All three modes CORRECT. The Mastery was banked as a RANKED (Hall-of-Fame) attempt.

ModeResultGM flagRanked
single_solveCORRECT (1.9s)Token094bb8647c20c86aa575cb97214d7cd1no
time_trialCORRECT (per-field diagnosis)Token276c5d9346ebedbc9c1123f4da1adeb3no
masteryCORRECT (7.7s)Tokenb4939eae932dd0b5b3b56e10758bb4dcyes

Prompt

Application databases have been copied, truncated, embedded, and mislabeled. Validate page geometry, connect databases to parent applications, and find the impossible claim.

Single Solve

One answer id: d23-sqlite-001-sqlite-page-size-bytes. The staged database-001.sqlite reads 4096 at offset 16-17 (u16 BE page size). AUTOSOLVE submitted 4096 in 1.9 s.

Time Trial

28 answer ids repeating every lane field so far (D04 jpeg, D05 png, D06 pdf, D07 zip, D09 pe, D12 dotnet, D13 msi, D14 lnk, D15 ooxml, D16 email, D17 audio, D18 video, D19 elf, D20 macho, D21 apk, D22 jar) plus the Day 23 sqlite field and a sha256 synthesis. The first autosolve pass was rejected as a batch with no per-field feedback. The GM's op=submit route accepts partial batches and returns {"result":"accepted","accepted_ids":[...],"remaining":N} per field, so single-field submits isolated the two wrong extractors:

  1. d07-zip-001-zip-central-directory-member-names - rejected in central-directory order; accepted when the array is sorted lexically. (OOXML's external_relationship_targets already sorted since Day 15; Day 23 extended that convention to zip CD names.)
  2. d20-macho-001-macho-cpu-type-set - rejected in file order; accepted when the unique cputypes are sorted ascending as 8-digit hex: ["0x00000007","0x00000012","0x01000007"] (i386, PPC, x86_64 w/ABI64).

Every other field was accepted on the first per-field submit (including PE ARMNT 0x01c4, ELF 64-bit header with EM_PPC 0x0014, dotnet MVID in LE reading [id], MSI product_code with braces upper {[id]}, SQLite page_size 1024, synthesis order ["video-001.mp4","panel-001.png","evidence-note.txt","database-001.sqlite","installer-001.msi"]).

Mastery

Multi-set bundle (set1..set6 of the same question shape). After the Time Trial exposed the two format bugs, both extractors were patched - zip.central_directory_member_names sorted, macho.cpu_type_set sorted set - and the Mastery autosolve ran once: ranked, CORRECT in 7.7 s.

Fixes landed in the framework

  • challenges/_lib/autosolve_metadata.py::zip_central_directory_member_names now returns sorted(names) instead of CD scan order. The GM switched to sorted on Day 23; CD-scan order was rejected for every batch.
  • challenges/_lib/hdr_facts.py::macho_facts now returns jarr(sorted(set(cpus))) instead of jarr(dict.fromkeys(cpus)) for cpu_type_set on fat Mach-Os. 8-digit lowercase hex sorts numerically.

Lessons

  • op=submit accepts partial batches. On a batch rejected without feedback, bisect by submitting each answer id alone - the response names accepted_ids and remaining, so one wrong field is pinpointed in N free reads. Set HUNTRESS_SUBMIT_CAP=0 so the client-side cap does not refuse the probe sweep.
  • Day 23 moved the macho and zip collections to the same "sorted set" convention that OOXML has used since Day 15. Treat every multi-value field in metadata-forensics as sorted by default from Day 23 on.
  • The challenge's "find the impossible claim" wording did NOT point at a header-vs-reality lie that we had to correct. It referred to the format convention mismatch: our file-order answers were the "impossible claim" (an unsorted set).