Selective Memory (Day 14, binary-reverse-engineering)
- GM id:
binary-reverse-engineering-day-14-objective-c-persistence-descriptor - Lane: binary-reverse-engineering | Agent: agent-2 (homestation)
- Modes: single_solve CORRECT, time_trial CORRECT, mastery CORRECT (all three in one banked AUTOSOLVE pass)
- Artifact:
challenges/Day 14/_stage/Selective_Memory/artifact_0.bin(32,246,719 bytes, sha256cd745ce710566bf4...), AES zip, passwordinfected, 15 members - Fields: 140 (single/time_trial), 420 (mastery = 3 sets)
Flags (GM-accepted)
| Mode | Flag | Time |
|---|---|---|
| single_solve | Token9d185d00fb1174989880389c6efbd07d | 3.1 s |
| time_trial | Tokene6cea396710c51585ed115f66b8dadc9 | 9.4 s |
| mastery | Tokend94fe227e767da1179b2bbe543ad2835 | 48.6 s |
Solved under handle the team solver; all three submits returned 200 CORRECT on the first attempt (no ranked/run_for_time entry was used, per standing policy).
The challenge
GM prompt: "Objective-C never forgets: every class, every selector, every string is registered and waiting. Read the persistence descriptor and learn what this binary remembers."
This is the same cumulative multi-field forensic inventory the lane has asked every day since Day 1. The evidence zip bundles one specimen per earlier day (d01 .. d14) and asks the same ~17 answer shapes over each: ELF/PE header facts, container and architecture labels, section and symbol tables, a recover_code disassembly, JAR compression methods, .NET metadata version, and the embedded single-byte-XOR HCFG config record.
Day 14's new layer is the toolchain only: an Objective-C (GNUstep / libobjc2 runtime, GNU gold linker) native x86-64 ELF.
Day 14 member
d14--variant-000.elf (challenge artifact, not published), 17,304 bytes, \x7fELF ELF64 LSB, 38 sections. The Objective-C runtime registration tables the prompt alludes to are plain ELF sections, GNUstep-style (not the Apple __objc_classlist / __objc_methname Mach-O spelling, which is why an Apple-flavoured grep finds nothing here):
__objc_selectors __objc_class_refs __objc_classes __objc_constant_string
__objc_cats __objc_protocols __objc_protocol_refs __objc_class_aliases
Symbols confirm one user class and its registered selectors:
- class:
_OBJC_CLASS_TriggerDescriptor(+_OBJC_METACLASS_,_OBJC_INIT_CLASS_,_OBJC_REF_CLASS_), baseNSObject - ivar:
__objc_ivar_offset_TriggerDescriptor._triggerName. - selectors:
triggerName,initWithTrigger:,eventCountForModel:(typed_i24), plus runtimealloc,copy,init - constant string:
objc_str_UserLogin(anNSConstantStringin__objc_constant_string) - runtime glue:
objc_init,objc_ctor,objcv2_load_function,__objc_load,objc_msgSend,objc_autoreleasePoolPush
The persistence descriptor (.hcfg)
Unchanged family mechanic: a .hcfg section holding HCFG + version byte + XOR key + CRC32, then a single-byte-XOR JSON blob.
- section
.hcfg, record at file offset 4000 - header:
48 43 46 4701(version 1)a1(XOR key0xa1)6f 01da 91 f4 79(crc32 LE =0x79f491da) - decoded payload:
{"campaign":"BR14-000","domain":"node-000.day14.example.invalid","ipv4":"198.51.100.20",
"url":"https://node-000.day14.example.invalid/api/BR14-000","mutex":"Global\\HuntressBR14_000", ...}
Submitted d14_* answers (10 ids, all reusing existing slot shapes):
| id | value |
|---|---|
d14_artifact_sha256_1 | d93c0e4d52fc9111b9133ca56b8da3122c9b4186f967102c3bb5b0fe9fd3b617 |
d14_analysis_format | elf |
d14_analysis_architecture | x86-64 |
d14_analysis_machine | x86-64 |
d14_analysis_config_campaign | BR14-000 |
d14_analysis_config_domain | node-000.day14.example.invalid |
d14_analysis_config_url | https://node-000.day14.example.invalid/api/BR14-000 |
d14_analysis_config_mutex | Global\HuntressBR14_000 |
d14_analysis_config_crc32_hex | 0x79f491da |
d14_analysis_has_hcfg_section | true |
Why it was a zero-work day
challenges/_lib/autosolve_binary.py (plus binary_facts / hdr_struct / config_facts) already covers every slot shape in this contract. The prebrief contract diff said it outright: single/time_trial shared 130, changed 0, new 10; mastery shared 390, changed 0, new 30 - i.e. the d14 block is ten copies of shapes the adapter has answered since Day 2, so zero new slots (the fourth day in a row: D11 D/LDC, D12 Pascal, D13 Crystal, D14 Objective-C). The readiness gate reported 140/140 and 420/420 ids resolved on the Single stand-in, so AUTOSOLVE opened the timed windows itself:
- time_trial: EARLY slot, rehearsed finish T+16.9 s <= T+18 s -> started immediately, CORRECT at 9.4 s
- mastery: rehearsed finish was past T+20 s, so the LATE slot rule held it to T+60 s with nothing spent, then CORRECT at 48.6 s
No adapter extension, no new extractor, no .needuser. The novelty words (descriptor, selector, registered, remembers) pointed at the runtime tables, which turned out to be flavour rather than a new question.
Reusable notes
- GNUstep Objective-C on Linux registers via
__objc_selectors/__objc_classes/__objc_constant_stringsections andobjcv2_load_function; Apple section names (__objc_methname,_OBJC_CLASS_$_) do not appear - classify by the GNUstep spelling when the container is ELF. - The
.hcfgXOR key is in the record header byte 5 (here0xa1); it does not need guessing, and the CRC32 at bytes 8..11 is little-endian. - Grader label conventions held again:
elf,x86-64, lowercase0xhex,true/falseas strings, base-10 integers.
Files
scripts/inspect_d14.py- extracts the AES zip, lists members, fingerprints the d14 ELF (sections, ObjC markers, HCFG offset)scripts/decode_hcfg.py- recovers the XOR key and decodes the.hcfgJSON, lists the ObjC class/selector symbolsscripts/wu_extract.py- pulls the GM op/answer record for this writeup