Tactical LogHuntress CTF 2026 // team archive

Days / Day 09 / Binary RE / Nimble

DAY 09//LANE: BINARY RE//STATUS: CORRECT x3

Nimble

Binary RE SSTTMM Modes reported CORRECT in this write-up

Day 09 - Nimble (binary-reverse-engineering) - SOLVED 3/3

GM id: binary-reverse-engineering-day-09-nim-generated-runtime Family: dual_format_re (the cumulative Day 01-09 forensic-inventory series) Agent: agent-4

modefieldsresultflag
single_solve90CORRECTToken7c271154aae3492ed2cb844ef3a62e9f
time_trial90CORRECT (4.5s)Token5c4fcc7c6a4c5efedc50f3878fa1abdf
mastery450 (5 sets)CORRECT (17.1s)Tokenef3687ea7c29b42da1db19a6bc84080b

What the day added

The delivery is the usual cumulative evidence.zip (challenge artifact, not published) of ten members, one per earlier day plus today's:

d01--variant-NNN.elf   C ELF              d06--variant-NNN.jar   Java
d02--variant-NNN.elf   C++ ELF (HCFG)     d07--variant-NNN.jar   Kotlin
d03--variant-NNN.dll   PE/.NET (+ .runtimeconfig.json)
d04--variant-NNN.elf   Go                 d08--variant-NNN.elf   Swift (67 MB)
d05--variant-NNN.elf   Rust               d09--variant-NNN.elf   Nim   <-- new

d09 is a Nim program compiled through C to a 60 KB x86-64 ELF. The banked challenges/_lib/autosolve_binary.py already answered 89 of the 90 fields on the first claim, so the work was three specific gaps - all now folded back into the adapter.

1. d09_analysis_hcfg_offset (new field, no precedent anywhere in the repo)

Every earlier day asked hcfg_version; Nim day asks for an offset instead, as a base-10 integer. The Nim runtime does not carry a raw HCFG record: _hcfg_scan finds only two decoys, both Nim artefacts in .rodata,

30633  "v000.nim(13, 3) `raw[0 .. 3] == \"HCFG\"` "   <- Nim range-check message
30664  "HCFG"                                        <- the comparison literal

The real record is base64 text held as a Nim const string (length-prefixed: fc 01 = 508 bytes, then the 0x40... flags word, then the data):

file offset 30696: SENGRwFRcAGo4RSW...   -> 48 43 46 47 01 51 70 01 a8 e1 14 96 ...
                                            H  C  F  G ver key  len    crc32

xor key 0x51, hcfg version 1, crc32 0x9614e1a8, campaign BR09-000.

The graded answer is the file offset of the base64 text itself: 30696. Accepted first try. Ruled out by that acceptance: the virtual address (0x4077e8 = 4225000), 0 and 12 (positions inside the decoded record), and 2024 (offset within .rodata).

2. d09_analysis_machine is integer-kind

Day 01-07 grade analysis_machine as the text x86-64; Day 08 and Day 09 grade it as the decimal e_machine (62), while d09_analysis_architecture stays text x86-64. The adapter had a hardcoded d08_ prefix test for this. It now reads each question's own answer_format.kind from the start payload, so a future day flipping a field's kind needs no code change; the d08_/d09_ prefix list survives only as the offline fallback for a task file with no questions block.

3. d01_analysis_disassembly on a stripped variant (this is what cost the first time trial)

The first time-trial attempt banked 89 of 90 and left one wrong. The mutated delivery's d01--variant-002.elf (challenge artifact, not published) is stripped - no .symtab, so no recover_code FUNC symbol - and the question says plainly: "If that symbol is absent, submit an empty array." The adapter was falling back to a .text scan for a mov eax, imm32; ret stub and submitting that guess. This is the exact rejection the KB had already recorded from the Day 01 time trial (BR01-STRIPPED-PROFILE, immediate 4099 rejected) - the cause was never the immediate's value, it was answering at all.

disasm_recover() now requires the FUNC symbol in the symbol table and returns [] when it is absent. When the symbol is present it disassembles that symbol's own bytes, keeping the proven rendering for the canonical stub (["mov eax, 8100","ret"] - decimal immediate, no addresses or byte columns) and otherwise rendering capstone output with standalone hex immediates converted to decimal.

Verified offline against both deliveries before any timed window: single-solve variant-000 -> ["mov eax, 8100","ret"], time-trial variant-002 -> [], 90 sure / 0 probe / 0 unknown on each, and _test_autosolve_binary.py --offline-d12 still PASS (no Day 01/02 regression).

Method note

Single Solve is untimed, so it was the lab: the two open fields were probed one id per submit (scripts/probe_d09.py + scripts/cands.json) with the candidate list ordered by what the bytes actually supported. Both landed on the first candidate. Only after the adapter reproduced 100% of both the single-solve and the recovered time-trial variant did a timed window get opened, and each timed mode was then a single automated shot.

Files

  • challenges/_lib/autosolve_binary.py - extended (3 changes above); this is the artifact that matters for Day 10+
  • scripts/diag.py - offline replay: member inventory + sure/probe/unknown for a delivery
  • scripts/probe_d09.py, scripts/cands.json - one-id-per-submit prober used on the untimed mode
  • failed-attempt bytes kept by the adapter at [home path]<id>_time_trial/ - this is what made the stripped-variant diagnosis possible after the window closed