elf aware: headers, symbols, sections, and a return constant
huntress ctf 2026, day 1 · binary reverse engineering · all three modes
tools used: binary-analysis toolkit and timed-attempt pipeline.
the archive held a small, unstripped c elf and asked for nine properties. this was less about decompilation than about knowing which layer owns each fact: identification bytes, the main header, section and symbol tables, a custom section, and two instructions of disassembly.
the key insight: most fields are fixed offsets in the header, and the hard part is the exact vocabulary each field expects.
artifact anatomy
the single sample was elf64, little-endian, x86-64, ET_EXEC, with an entry point (0x401030). a .symtab section existed, a custom .huntress string section held a marker (BR01-SYMBOL-ORIENTATION), and one custom global function (recover_code) was the exported symbol the questions asked about.
e_ident offset 4 is the class and offset 5 the byte order. e_type and e_machine are 16-bit values at offsets 16 and 18; e_entry is a 64-bit value at offset 24 for elf64.
extraction method
readelf -hW symbol-orientation.elf
readelf -SW symbol-orientation.elf
readelf -sW symbol-orientation.elf
readelf -p .huntress symbol-orientation.elf
objdump -d symbol-orientation.elf
the parser reads the header itself, using the declared endianness rather than assuming x86:
little = blob[5] == 1
endian = "<" if little else ">"
e_type = struct.unpack_from(endian + "H", blob, 16)[0]
e_machine = struct.unpack_from(endian + "H", blob, 18)[0]
e_entry = struct.unpack_from(endian + "Q", blob, 24)[0]
the exported symbol is a defined global FUNC after removing compiler and runtime names such as _start, main and initialization helpers, which is more robust than the first symbol in nm output. its body disassembles to a move-immediate and ret; the solver scopes its regex to that function and extracts the immediate.
validation and answer formatting
the nine ids were elf_class, endianness, machine, elf_type, marker, has_symtab, exported_symbol, the return-constant field (named after the function whose constant it asks for) and entry_point. most were exact_text ("exact artifact-derived value only ... use the literal none only when the requested field is absent"); has_symtab was an enum of true and false, entry_point a pattern ("lowercase hexadecimal with a 0x prefix"), and the return constant an integer ("base-10 integer with no units, separators, or explanatory text") on the single.
each field follows its own convention: ELF64 (readelf's spelling), x86-64 (the file command's), little, ET_EXEC (the spec constant name), lowercase true/false, the entry point as lowercase 0x hex, and the return value as a decimal string even though objdump prints the immediate in hex. a missing symbol answers none.
failed approaches and pivots
extraction was quick; vocabulary was not. the time trial did not land at first because the return constant's text form was unresolved between hex, decimal and none: in the timed sets that field came either as an integer or as exact_text with the none rule, set by set. the 0x hex form was rejected; decimal was right under both kinds, and none only when the binary has no function of the name the question gives. per-field probes on the untimed specimen settled every convention, and the solver encoded them so timed variants needed no guessing.
automation across modes
mastery rotated its layout on every start: sets times slots (slotM_ field prefixes against slot-M.elf members) times nine fields. the same per-binary function ran for every slot, and one start per attempt kept the material stable.
lessons
- know which structure owns a fact before reaching for a decompiler.
- record the grader's vocabulary per field; it differs even within one binary.
solve times
public scoreboard · day 1 finish: below top 10
- 2h43msingle#63 of 434 solvers
- 3h03mtime trial#62 of 392 solvers
- 3h04mmastery#58 of 383 solvers
- #54team day finishof 358 teams that cleared it, in 3h04m
- #3single orderof the day's 5, by our flag time
from the public scoreboard: 09:00 et release to our flag; the table below counts from our attempt's start.
from our first start of the mode to its flag; earlier preparation was not recorded.
| mode | work to flag | attempts | wrong submits |
|---|---|---|---|
| single | 14m36s | 1 | - |
| time trial | 2s on the clock (prep not logged) | 1 logged (earlier tries not recorded) | not logged |
| mastery | not recorded | 1 | - |
the clock
server timer of the winning attempt; used is attempt start to flag, left is what remained on the clock. day 1-2 deadlines were not logged per attempt.
| mode | time available | time used | time left |
|---|---|---|---|
| time trial | not recorded | 1.6s | not recorded |
| mastery | not recorded | not recorded | not recorded |
where we got stuck
- the time trial did not land on the first tries: the return-constant field's exact text form (hex, decimal, or
none) was unresolved until probing showed decimal. - mastery rotated its layout on every start (sets x slots x nine fields), so starting more than once per attempt threw the material away.
flags: single, time trial, and mastery [redacted].
{generated by a clanker}