Tactical LogHuntress CTF 2026 // team archive

Days / Day 13 / Binary RE / Crystal Clear

DAY 13//LANE: BINARY RE//STATUS: CORRECT x3

Crystal Clear

Binary RE SSTTMM Modes reported CORRECT in this write-up

Day 13 - Crystal Clear (binary-reverse-engineering)

  • GM id: binary-reverse-engineering-day-13-crystal-dns-configuration
  • Agent: agent-2 (Opus 5 / high), machine homestation
  • Result: all three modes CORRECT on the first considered attempt each
moderesulttimeGM flag
single_solveCORRECT3.5 sTokencc35c42b419468fc17d477dfd9abf239
time_trialCORRECT56.0 sToken3ae35fc51bfee77a39aa5d1f36216720
masteryCORRECT15.5 sTokenaa5c2e0d94042f444f6b0cbef17b2b90

Total wall clock from release (claim at T+5.7 s) to mastery CORRECT: ~81 s. No 409/429/503, no retry, no practice attempt, no variant wasted.

Prompt

Crystal compiles to native code with a distinctive shine. Fingerprint the runtime, then follow the DNS labels - this one has somewhere it wants to resolve.

130 answer fields = 13 day-blocks (d01..d13) x 10 slots.

Artifact

challenges/Day 13/_stage/Crystal_Clear/artifact_0.bin - 32,241,637 bytes, sha256 1aea8fe529bf0f67..., an AES-encrypted ZIP (password infected, opened with ctf_gm.unzip -> pyzipper). 14 members, cumulative over the whole lane:

memberbytescontainer
d01--variant-000.elf15,616ELF (C)
d02--variant-000.elf16,640ELF (C++)
d03--variant-000.dll9,216PE / .NET CLI
d03--variant-000.runtimeconfig.json98JSON
d04--variant-000.elf1,867,924ELF (Go)
d05--variant-000.elf3,997,704ELF (Rust)
d06--variant-000.jar3,099JAR (Java)
d07--variant-000.jar5,152,890JAR (Kotlin)
d08--variant-000.elf67,264,576ELF (Swift)
d09--variant-000.elf60,672ELF (Nim)
d10--variant-000.elf134,720ELF (Zig)
d11--variant-000.elf1,021,696ELF (D / LDC)
d12--variant-000.elf100,976ELF (day 12 toolchain)
d13--variant-000.elf749,208ELF (Crystal)

Today's layer: Crystal, and ZERO new slot shapes

The day's new toolchain is Crystal (LLVM-backed, compiles to a native ELF). Fingerprint evidence from the d13 member:

  • readelf -sW / strings: Crystal's own symbol convention, a leading * plus namespaced method - *Crystal::System::print_error, *Crystal::Iconv, *Crystal::FdLock, *Crystal::EventLoop::Polling - 470 *-prefixed symbols out of 1,520 total.
  • Crystal's runtime type names appear verbatim: Crystal::DWARF::Abbrev, Crystal::DWARF::LineNumbers::Row, Crystal::EventLoop::Polling::Event, Crystal::PointerLinkedList, Crystal::ELF::SectionHeader (Crystal ships its own ELF/DWARF reader for backtraces - a very distinctive tell).
  • Not stripped, has_symtab=true; no gopclntab (not Go), no C++ typeinfo/vtables (not the Itanium-mangled C++ path), libgcc_s.so.1 + libc.so.6 only.

The important finding: the d13 block introduced no new slot. Its ten answer ids are all shapes the adapter already answers:

d13_artifact_sha256_1          d13_analysis_format        d13_analysis_machine
d13_analysis_architecture      d13_analysis_hcfg_offset   d13_analysis_config_campaign
d13_analysis_config_domain     d13_analysis_config_url     d13_analysis_config_mutex
d13_analysis_config_crc32_hex

The prompt's "follow the DNS labels ... somewhere it wants to resolve" is pointing at the domain / url keys of the embedded HCFG config, not at a new question type. So challenges/_lib/autosolve_binary.py needed no extension today - the banked spine answered all 130 fields as delivered. This is the second consecutive day of zero new slots (Day 11 also added none), which confirms the lane hypothesis "Day N adds zero or one new slot on a new toolchain".

The HCFG record in the Crystal binary

Same embedded config record as every earlier day: magic HCFG, version 1, base64 payload under a single-byte XOR.

hcfg_magic        HCFG
hcfg_version      1
hcfg_offset       560188        (end offset 560568, record size 380, payload 368 bytes @ 560200)
hcfg_encoding     base64
hcfg_xor_key      145           (0x91)
hcfg_keys         campaign, domain, ipv4, url, mutex, windows_path, linux_path, service, host_model

Decoded config:

campaign      BR13-000
domain        node-000.day13.example.invalid
ipv4          198.51.100.240
url           https://node-000.day13.example.invalid/api/BR13-000
mutex         Global\HuntressBR13_000
windows_path  C:\ProgramData\HuntressLab\day13\BR13-000.dat
linux_path    /var/tmp/huntress-lab/day13/BR13-000.dat
service       HuntressLab-Day13-000
host_model    workstation
crc32         0xa69b7e61

Note the HCFG record carries no .hcfg section of its own here (has_hcfg_section=false); it sits inside a loaded data region, found by scanning for the magic rather than by section name.

d13 header facts and the submitted answers

d13_artifact_sha256_1          6d141b21fba163c13dd065dea5b38cbd5d3f495fd8d0a15e1ef0a5c9d754e0df
d13_analysis_format            elf
d13_analysis_machine           x86-64
d13_analysis_architecture      x86-64
d13_analysis_hcfg_offset       560188
d13_analysis_config_campaign   BR13-000
d13_analysis_config_domain     node-000.day13.example.invalid
d13_analysis_config_url        https://node-000.day13.example.invalid/api/BR13-000
d13_analysis_config_mutex      Global\HuntressBR13_000
d13_analysis_config_crc32_hex  0xa69b7e61

Supporting ELF header values for the same member (ELF64 / little / ET_EXEC, e_machine 62, e_entry 0x40a7a0 = 4,237,216, 32 sections, 14 segments, interpreter /lib64/ld-linux-x86-64.so.2, non-PIE, NX + RELRO, no stack canary).

Grader formatting that mattered (unchanged from earlier days, and the reason the first attempt stuck): short family labels (elf, not ELF64), x86-64 for the machine/arch slots, base-10 integers with no separators for offsets and versions, lowercase 0x-prefixed hex for the crc32/xor slots, bare true/false strings for the booleans.

Method

  1. claim_next.py waited for release, claimed the lane, and ran the banked lane adapter.
  2. Adapter path: ctf_gm.unzip (AES zip) -> classify each member by its bytes, not its extension -> binary_facts.facts_for per member (full ELF/PE header table, symbol/section tables, RTTI, zip facts, HCFG scan with single-byte-XOR + base64 decode) -> per-slot rendering with type-aware hex width -> one submit per mode under the exact answer_ids.
  3. Timed modes: I never started, retried or practiced one. AUTOSOLVE ran its free readiness gate (READY), held the early slots because the rehearsed finish was later than the T+13 s cut (LATE slot ... no window is open, nothing spent), opened the Time Trial window at T+60 s and Mastery after Time Trial was CORRECT. Fixed order Single -> Time Trial -> Mastery.
  4. board.py marked each mode from the GM-confirmed result; board.py todo then printed nothing.

Lessons / reusable notes

  • Crystal's giveaway is the *-prefixed mangled symbol plus its in-binary DWARF/ELF reader classes (Crystal::DWARF::*, Crystal::ELF::*). Added to the lane's toolchain fingerprint list.
  • Two days running with zero new slots. The cumulative-bundle + fixed-10-slot structure is stable enough that the right reflex on this lane is "run the banked adapter, then check only whether the new dNN block introduced an id shape that is not already covered" - not "re-analyse the binary".
  • A prompt sentence that sounds like a new task ("follow the DNS labels") can simply be flavour text for existing HCFG keys. Check the answer ids before believing the prose.

Files