Day 15 - Stripped (binary-reverse-engineering)
GM id: binary-reverse-engineering-day-15-assembly-and-symbol-loss Lane: binary-reverse-engineering | Agent: agent-2 | Date: 2026-10-15
Result
| Mode | Result | Flag |
|---|---|---|
| Single Solve | CORRECT | Token4c3d3b45b970c89e04e117194504392d |
| Time Trial | CORRECT | Tokenb4da6075cfc40804d87dd254cb6b9ebb |
| Mastery | CORRECT | Token4ae80f57912498fc9ac9e1b5475b0509 |
All three modes landed on the first submit each (single 5.2 s, time trial 6.3 s, mastery 17.3 s), with 150 / 150 / 750 fields resolved locally and zero probe or unknown fields.
Prompt
No symbols. No names. Just raw assembly and the calling conventions it can't hide. Reconstruct the stripped functions from the instructions alone - the stack never lies.
Artifact
artifact_0.bin (challenge artifact, not published) is the usual AES-256 evidence zip (password infected, 32,251,967 bytes, sha256 a297e8326c70e79c...). It is the cumulative Day 01-15 bundle, but the Day 15 deliverable is the first one that is a pair of binaries instead of a single file:
d15--variant-000-symbolized.elf 16160 bytes sha256 83a58b2e4162ecd25360cd898ec7eb57b5e1fe820aa0c28c5df774176e104686
d15--variant-000-stripped.elf 14360 bytes sha256 b97e0d28b11036b69552c01d42107411b7fa55fd371d6797027f3c5742dc9c7e
Both are the same build: ELF64, little endian, ET_EXEC, e_machine 62 (x86-64), entry 0x401060, identical .text (0x401060, 0x18d bytes) and an identical .hcfg section at file offset 0x210d. The stripped copy is the symbolized copy with its symbol table removed - 28 section headers instead of 30 (.symtab / .strtab gone), 14,360 bytes instead of 16,160.
What was new today
The title hint ("assembly and symbol loss") pointed at disassembly work, but the actual novelty was in the question ids, not the analysis. Day 15 added no new fact - it added a new addressing shape: the member name moved into the field id, so an ordinary inventory question now names which half of the pair it asks about.
New question keys (8), plus one that had never fired on a day with no aux json:
d15_analysis_stripped_format d15_analysis_symbolized_format
d15_analysis_stripped_machine d15_analysis_symbolized_machine
d15_analysis_stripped_architecture d15_analysis_symbolized_architecture
d15_analysis_stripped_config_campaign
d15_analysis_stripped_config_domain
d15_artifact_sha256_2
The banked adapter resolved 141 / 150 and correctly HELD the mode (nothing submitted): it parses dNN_<suffix> and had no member selector, so analysis_stripped_machine was not a suffix it knew, and artifact_sha256_2 only ever meant "the dNN's aux .json" (the Day 03 .runtimeconfig.json), which Day 15 does not have.
Fix
Two changes in challenges/_lib/autosolve_binary.py - member routing, not a new extractor:
_variant_tags(files, dk)collects{tag: bytes}for deliverables nameddNN--variant-XXX-<tag>.<ext>. Inanswer(), a suffix matchinganalysis_<tag>_<rest>- where<tag>is a tag actually present in a filename - rebindsmainto that member and rewrites the suffix to the plainanalysis_<rest>. Every existing handler then answers it unchanged. Gating on tags that really exist is what keeps plain suffixes (analysis_config_*,analysis_has_*,analysis_elf_*,analysis_xor_key_hex) from being mis-split.artifact_sha256_2falls back to the dNN's second non-json member in evidence-zip order when there is no aux json.
The ordering of artifact_sha256_1 / _2 was the only genuinely ambiguous call. The evidence zip is alphabetical everywhere except Day 15, where it writes -symbolized.elf (challenge artifact, not published) before -stripped.elf (challenge artifact, not published) (stripped < symbolized alphabetically, so this is not a sort). That member order is the generator's own artifact list, and it is the same order the GM-graded Day 03 pair uses (.dll then .runtimeconfig.json). So _1 = symbolized, _2 = stripped - confirmed by the GM accepting the batch.
d15 answers
| field | value |
|---|---|
d15_artifact_sha256_1 | 83a58b2e4162ecd25360cd898ec7eb57b5e1fe820aa0c28c5df774176e104686 |
d15_artifact_sha256_2 | b97e0d28b11036b69552c01d42107411b7fa55fd371d6797027f3c5742dc9c7e |
d15_analysis_symbolized_format | elf |
d15_analysis_symbolized_machine | 62 |
d15_analysis_symbolized_architecture | x86-64 |
d15_analysis_stripped_format | elf |
d15_analysis_stripped_machine | 62 |
d15_analysis_stripped_architecture | x86-64 |
d15_analysis_stripped_config_campaign | BR15-000 |
d15_analysis_stripped_config_domain | node-000.day15.example.invalid |
.hcfg (single-byte XOR key 0xb1, HCFG version 1, offset 8461, CRC32 0xc348914e) decodes identically in both halves - stripping symbols does not touch the config record:
{"campaign": "BR15-000", "domain": "node-000.day15.example.invalid", "host_model": "workstation",
"ipv4": "198.51.100.40", "linux_path": "/var/tmp/huntress-lab/day15/BR15-000.dat",
"mutex": "Global\\HuntressBR15_000", "service": "HuntressLab-Day15-000",
"url": "https://node-000.day15.example.invalid/api/BR15-000",
"windows_path": "C:\\ProgramData\\HuntressLab\\day15\\BR15-000.dat"}
Reproduce
wsl python3 'challenges/Day 15/Stripped/scripts/check_d15.py' # offline: 150 sure, 0 probe, 0 unknown
wsl python3 challenges/_lib/autosolve.py 15 binary-reverse-engineering-day-15-assembly-and-symbol-loss binary-reverse-engineering
Lesson for the lane
A dNN is no longer guaranteed to be one binary. When a day's member name carries a build variant, the variant lands in the field id, so the adapter needs a member selector, not a new fact extractor - and artifact_sha256_N is numbered from evidence-zip member order, not from "main plus aux json".