Tactical LogHuntress CTF 2026 // team archive

Days / Day 31 / Binary RE / Final Boss

DAY 31//LANE: BINARY RE//STATUS: CORRECT x3

Final Boss

Binary RE SSTTMM Modes reported CORRECT in this write-up

Final Boss - Day 31 (binary-reverse-engineering-day-31-reversing-gauntlet)

Agent: agent-2. Lane: binary-reverse-engineering. Result: all three modes CORRECT.

ModeResultFieldsFlag
Single SolveCORRECT (untimed)4347a6cc4fe92e4ead27e48b9680404432f
Time TrialCORRECT, ranked, 7.6 s434778f4107f6cf04478daa7f28aa51a273
MasteryCORRECT, ranked, 69.7 s6510 (15 deliveries)647cf0f593625a4453122594449e511f

Challenge shape

"The gauntlet: everything you've learned ... in one campaign." The artifact is an AES-256 evidence.zip (password infected, 46 members). It has one dNN-- member group for each of Days 01-31 (ELF, PE, .NET, JAR, Neko .n, and others). The questions are the cumulative per-day slot set (dNN_analysis_*, dNN_artifact_*, dNN_structural_*). Day 31 adds a single new record: HGAU.

What happened

  1. On the first autosolve pass, the fresh D31 clone (a copy of the shared adapter) answered 390 of 434 fields. It held the other 44. Nothing wrong was sent; the 390 were banked.
  2. 38 of the 44 held fields were Day 26-30 slots. The shared adapter never got them, but the agent-2 Day 30 clone (Day 30/⚙️ Campaign Trail/scripts/binary_wip.py) already had all of them:
    • structural_artifact_count and structural_ordered_set_sha256
    • cli_runtime and has_decoy_branch (UTF-16LE at both byte alignments)
    • class_count, class_name, method_names_sha256, crash_fault_method, crash_frames_match_methods
    • has_nim_runtime, has_fpc_runtime, has_gnat_runtime
    • d_symbol_count and crystal_symbol_count (.symtab only)
    • conditional_branch_opcode_count
    • handler_symbol_count and handler_symbols_sha256
    • packing_key_1_hex and packing_key_2_hex (PK19 overlay k1 and k2)
    • the d29 optimized / unoptimized member aliases (O3 / O0)
    • the d30 HGRF capsule and the neko_bytecode format
  3. I ran both clones offline with the TT question set on single_0.bin (scripts/compare_ports.py). They disagreed on 0 of the 390 sure answers, nothing was lost, and the D30 clone answered 38 more.
  4. New for D31 is the HGAU record, found by a backward scan of the d31--sample.elf (challenge artifact, not published) member (scripts/hgau_probe.py). It uses the same 12-byte layout as HCFG: magic | ver | xor key | len u16le | crc32 u32le | payload.
    • There is exactly one CRC-valid record: offset 505640, ver 1, key 0xa1, len 1323, crc 0x5ed06133. It ends exactly at EOF.
    • The JSON schema is huntress.binary.reversing_gauntlet.v1.
    • The question's gauntlet. prefix is the record ROOT, the same convention as D30's graph.:
      • gauntlet.config.<k> reads config.<k>.
      • gauntlet.trace_branch reads trace.branch (success on the Single).
  5. I added _hgau_record() and an analysis_gauntlet_* dispatch to the clone. The record must be unique or the field is held, and trace_branch is checked against the stated vocabulary.
    • The offline result was 434 sure, 0 probe, 0 unknown.
    • The runner rerun then went Single Solve CORRECT, TT CORRECT (ranked), and Mastery CORRECT (ranked).

Lessons + D+1

  • What was missing: the shared autosolve_binary.py / fresh clone template lacked the D26-D30 extractors that each day's agent-2 clone already carried, plus the D31 HGAU record. The 38 D26-D30 fields were a porting gap, not new analysis.
  • Stall time: about 20 min from claim to all-CORRECT. Roughly 12 min of that went to working out the 44-field residue and finding the D30 clone. Diffing the two clones and adding HGAU took about 5 min, and the runner took about 2 min.
  • Port to shared, not clone-only:
    • the D26-D30 handler block (lines ~1107-1325 and ~2196-2258 of the D30 clone);
    • the _variant_tags O0/O3 alias;
    • the d30 HGRF block;
    • the new _hgau_record + analysis_gauntlet_* dispatch. The fresh-clone template for any later binary day should start from this Day 31 clone.
  • Generalisable rule: HCFG / HGRF / HGAU are one 12-byte record family with different magics. A future H??? record should go through one magic-parameterised scanner, with a uniqueness Hold.
  • Process: before patching a held residue, diff the previous day's per-CID clone against the fresh clone on the Single bytes. A zero-disagreement check on the already-accepted fields makes the swap safe in seconds.

Evidence retained: artifacts/ (single, time_trial, and 15 mastery deliveries); scripts/binary_wip.py (final); scripts/binary_wip.py.orig-d31 (rollback); compare_ports.py, hgau_probe.py, inspect_zip.py.