jni boundary: a native library, and an answer that was never across it
huntress ctf 2026, day 16 · mobile android · all three modes
tools used: android apk toolkit and timed-attempt pipeline.
the prompt: "identify a java-to-native type or length assumption that changes backend authorization." the apk kept day 15's shape: dex code, a challenge-owned native library, and a puzzle asset whose codec is named native.
the key insight: the story is about the boundary between java and native code, but the puzzle's data already reads as the answer's own shape, and the solver accepts a reading only in that shape. a codec name it does not know costs nothing when one of its readings validates.
challenge shape
single 1 field one apk
time trial 1 field another apk
mastery 5 fields five apks
the question names a case id and the app's package and asks for "the exact backend answer code returned for this case": ANS- and sixteen uppercase hex digits.
artifact anatomy
AndroidManifest.xml
classes.dex MainActivity, ChallengeActivity, JNIBoundary
assets/huntress_puzzle.json {"schema": ..., "day": 16, "codec": "native", "data": ...}
lib/x86_64/libarena.so the challenge's native library
META-INF/... the signature
the dex names the scenario (a jni_length key next to a "java-to-native validation mismatch" label) and an evidence endpoint that takes a case_id. the native library is a small elf that keeps its symbol table: two exported functions (a secret and the day number) and a read-only value in .rodata. in the single's apk that value and the puzzle's data are the same code.
the mastery's five apks carry earlier days' codecs, one per set: plain, hex, reversed, caesar with a key, and native again.
extraction method
the puzzle names its codec. a known name selects its transforms and key source; an unknown name selects none, so the data itself is checked first, and then the keyed and unkeyed searches run. the first reading in the answer's shape wins. for native the data is already an ANS- code.
def decode(puzzle, keys=None):
codec, data = codec_text(puzzle["codec"]), data_text(puzzle["data"])
try:
steps = steps_for(codec) # a known name: its transforms and key
except ValueError:
steps = [] # an unknown name: the data as it is
for out in chain(steps, data, puzzle):
if is_answer(out):
return out
keyed = keyed_peel(data) # single-byte keys, 2-4 byte keys
if keyed is not None: # recovered from the ANS- prefix
return keyed
text, path = peel(data, goal=is_answer) # every transform, breadth-first
return answer_in(text)
by hand, with nothing but the standard library:
import json, re, zipfile
apk = zipfile.ZipFile("jni_boundary.apk")
puzzle = json.loads(apk.read("assets/huntress_puzzle.json"))
assert puzzle["codec"] == "native"
assert re.fullmatch(r"ANS-[0-9A-F]{16}", puzzle["data"]) # the code, as stored
on the single, strings lib/x86_64/libarena.so | grep ANS- prints the library's .rodata copy of the same code.
validation and answer formatting
a reading is accepted only when it matches the question's pattern, ANS- plus sixteen uppercase hex digits; a reading that fails it is a crash, never a guess. the solver never reads the library; on the single its .rodata value and the puzzle's data agree, so either source would have done.
automation across modes
the same decoder answered the single, the time trial and all five mastery sets; each set's codec comes from its own puzzle asset.
lessons
- a challenge's narrative and its graded field can be far apart; read the field and its syntax before the story.
- a search fallback behind a name table absorbs a new name for free only when its output can be checked; here the answer's pattern is the check.
solve times
public scoreboard · day 16 finish: 1st
- 0:04.0singlefirst blood, +0.25 s
- 0:06.0time trialfirst blood, +0.66 s
- 0:08.8masteryfirst blood, +0.48 s
- #1team day finishof 162 teams that cleared it, in 1:02.1
- #1single 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 when we started working on the mode to its flag.
| mode | work to flag | attempts | wrong submits |
|---|---|---|---|
| single | 1.0s | 1 | 0 |
| time trial | 0.6s | 1 | 0 |
| mastery | 1.1s | 1 | 0 |
the clock
server timer of the winning attempt; used is attempt start to flag, left is what remained on the clock.
| mode | time available | time used | time left |
|---|---|---|---|
| time trial | 50.0s | 0.6s | 49.4s |
| mastery | 170.0s | 1.1s | 168.9s |
where we got stuck
- nowhere: every mode passed on its first attempt.
flags: single, time trial, and mastery [redacted].
{generated by a clanker}