Page Background Animation
3953 words
20 minutes
Securinets CTF Quals 2026 — Top 10 Finish & Challenge Writeups

Securinets CTF Quals 2026#

The official 36-hour online qualifier for Securinets CTF 2026 opened 17.10.2026 09:00 UTC. We finished inside the Top 10 and qualified for the finals.

EventSecurinets CTF Quals 2026
Format36-hour online qualifier
Opened17.10.2026 09:00 UTC
ResultTop 10 — qualified for the finals

Writeups for the two challenges covered here: SILENT TONE, an RF signals forensics problem disguised as a hardware bench dump, and RabbitHole, a nine-hop OSINT chain.


SILENT TONE — RF / Signals Forensics#

Three radios, one logic capture, and a pile of fake flags. The hardest challenge we cleared.

CAUTION

SILENT TONE plants seven fake flags across the zip comment, JSON metadata, USB logs and the RF itself — several written specifically to bait people (and LLM-assisted solvers) who grep for Securinets{ instead of demodulating. A string shaped like a flag is not evidence.

AttributeValue
ChallengeSILENT TONE
EventSecurinets CTF Quals 2026
CategoryRF / Signals Forensics
ProvidedA zip containing bench.pcap and capture.sal
ToolingPython, NumPy, SciPy, reedsolo, Matplotlib

TL;DR#

  • bench.pcap contains the transmitter’s boot banner, which documents the entire radio protocol.
  • capture.sal holds ~0.82 s of I/Q at 8 MS/s: a 71.3 kHz on/off tone plus three QPSK units — A, B and C — all transmitting on the same channel at once.
  • The tone carries UART text: Securinets{d3m0d_th3_j4mm3r_n0t_b34c0n}. It is a decoy, and the scoreboard rejected it.
  • Unit A is loud and sends plaintext telemetry (also a decoy). We demodulated it, re-encoded it, and subtracted it from the recording.
  • Unit C yields part 2 from single frames. Unit B yields part 1 only when frames 16–20 are combined coherently.
  • XOR part1 ‖ part2 with SHA-256(sorted UIDs ‖ counter) → the flag.

1. The Files#

SILENT_TONE.zip
├── bench.pcap # Linux usbmon capture from the bench
└── capture.sal # Saleae Logic 2 capture

capture.sal is just a zip archive:

Terminal window
$ unzip -l capture.sal
SDR_CAPTURE_V2 fc=433920000 fs=8000000 ref=Securinets{str1ngs_1s_n0t_dsp_try_h4rd3r}
digital-0.bin digital-1.bin analog-2.bin analog-3.bin trigger-store.bin meta.json

The zip comment hands us the centre frequency — 433.920 MHz — and the sample rate, 8 MS/s. It also hands us fake flag #1.

meta.json describes a capture that Logic 2 named beacon_433:

ChannelTypeName
0DigitalTX_EN
1DigitalREF_100K
2AnalogSDR_I
3AnalogSDR_Q (contains hidden characters — see the decoy table)

The device is a Logic Pro 8 sampling the analog channels at 8 MS/s for 0.8193 s. Someone wired an SDR’s I/Q outputs straight into a logic analyser.


2. The PCAP: The Transmitter Documents Itself#

bench.pcap is USB traffic. Extracting the bulk-IN payloads reveals three devices:

  • Device 7 — a Rigol DS1104Z oscilloscope answering SCPI commands (*IDN?, :WAV:MODE RAW, …). Irrelevant.
  • Device 8 — the same transmitter after a firmware update: PHANTOM-TX fw 2.4.0 … dfu: image ok … radio: config per release notes. This matters later.
  • Device 5 — the boot banner of one of the transmitters, which is effectively the protocol spec:
PHANTOM-TX fw 2.3.1 STM32F405 @ 168 MHz
uid : C1E20DC4 (crc8/maxim poly=0x31 refl init=0x00 -> 0x1c)
rf : 433.920 MHz QPSK dac_clk=8.000 MHz sym_div=64
pulse: rrc beta=0.35 ntaps=769
pre : zadoff-chu N=127 u=1+(crc8(uid)%126) -> 29 x[k]=exp(-j*pi*u*k*(k+1)/N)
frame: 127 pre + 16 uid syms (gray, unscrambled, msb-first) + 1020 data syms
burst: 80 frames
fec : rs n=255 k=223 gf=0x11d fcr=0 gen=2
scr : lfsr7 poly=0x91 seed=crc8(uid) -> 0x1c out=bit0 (xor on msb-first bitstream)
il : block 51x40 (write rows, read cols)
map : gray 00:+1+1 01:-1+1 11:-1-1 10:+1-1 (I<-b1 Q<-b0) /sqrt2
pld : part|len|data|crc32be|rnd -> 223 B ; part0 = plain, part k>0 = k-th slice of secret^ks
ks : sha256-ctr keyed by the paired-net uids
pa : xtr-7 rapp asat=1.55 p=2.0 ampm a=0.22 b=0.85 dpd=off
net : 3 units paired, same channel, tdma=off, relay=on

Three lines carry the whole challenge:

  • net : 3 units paired, same channel, tdma=off — expect three overlapping transmitters in time and frequency.
  • part0 = plain, part k>0 = k-th slice of secret^ks — plaintext parts are never the secret.
  • ks : sha256-ctr keyed by the paired-net uids — we will need all three UIDs.
NOTE

The banner is a spec sheet, not a reference implementation. It never says how the LFSR is wired, which direction the receiver must undo the interleaver, or what the other two UIDs are. Those came out of brute force.


3. Getting I/Q Out of the Saleae Capture#

Our first draft assumed the older Saleae analog layout — one header followed by float32 samples. The real files are version 3, type 101: a chunked layout of int16 samples.

def load_v3_analog(path):
b = open(path, 'rb').read()
assert b[:8] == b'<SALEAE>'
ver, typ = struct.unpack('<ii', b[8:16]); assert (ver, typ) == (3, 101)
fs = struct.unpack('<d', b[0x11:0x19])[0]
nchunks = struct.unpack('<Q', b[0x2b:0x33])[0]
out, pos = [], 0x33
for k in range(nchunks):
start, end, cnt = struct.unpack('<QQQ', b[pos:pos+24])
assert end - start == cnt, (k, start, end, cnt)
out.append(np.frombuffer(b, '<i2', cnt, pos+24))
pos += 1042732
return fs, np.concatenate(out).astype(np.float32)
# x = I + 1j*Q -> out/iq.npy

The banner gives 8 MHz / 64 = 125 kBd, so at 8 MS/s every symbol spans 64 samples. We strip DC using the quiet stretch before the bursts and decimate by 8, landing at 1 MS/s and 8 samples/symbol for everything that follows:

x = x - x[:300000].mean()
y = resample_poly(x, 1, 8) # 1 MS/s, sps = 8

4. What’s On The Air#

  • A narrow carrier at 71.3 kHz runs the whole capture. It is alone before the bursts (0–0.04 s) and after them (≈0.787 s onward).
  • A wideband QPSK burst spans ≈0.04–0.787 s, with a short dip near 0.41 s.

Raising the burst to the 4th power strips the QPSK modulation and leaves a spectral line at 4× the carrier offset. The strongest line puts the loud transmitter at ≈ +22.18 kHz:

X = np.abs(np.fft.fft(b**4, 1<<22)) # strongest line / 4 -> ~22.18 kHz

The two quiet transmitters do not show up at all yet.


5. Finding the Transmitters With Zadoff-Chu#

Each unit’s preamble is a Zadoff-Chu sequence whose root is derived from its own ID:

def crc8_maxim(data):
c = 0
for byte in data:
c ^= byte
for _ in range(8): c = (c >> 1) ^ 0x8C if c & 1 else c >> 1
return c
def zc(u, N=127):
k = np.arange(N); return np.exp(-1j*np.pi*u*k*(k+1)/N)
def shape(sym, sps=8): # RRC beta=0.35, span 12 symbols
up = np.zeros(len(sym)*sps, complex); up[::sps] = sym; return np.convolve(up, H)

For the banner UID, crc8_maxim(C1E20DC4) = 0x1c, so u = 1 + 28 mod 126 = 29.

The other two UIDs were unknown, so we searched. zcres.py correlates the recording against the RRC-shaped preamble for every root u = 1…126 across frequency offsets of −40 to +40 kHz in 1 kHz steps, keeping any root whose peak/median ratio clears 9. Three roots stood out — 29, 47, 104 — and all three repeat every 9304 samples at 1 MS/s, which is exactly one frame (1163 symbols × 8).

UnitUIDCRC8Root uOffsetRole
AC1E20DC40x1c29≈ +22 kHzloud, plaintext telemetry
BC0F4E5B20x2e47+4.08 kHzquiet, part 1
C145D75BF0x67104−9.25 kHzquiet, part 2

B’s and C’s UIDs are read off their own UID symbols later. Each one checks out: 1 + CRC8(UID) mod 126 reproduces the root we found by search.

Frame layout:

| 127-symbol ZC preamble | 16-symbol UID | 1020 data symbols | = 1163 symbols

6. Demodulating Unit A#

Per frame: shift to the unit’s offset, apply the RRC matched filter, then take every 8th sample starting at the correlation peak — with a +96 offset for the combined transmit/receive filter delay. Remove the leftover phase ramp with a linear fit over the preamble, then track phase with a decision-directed PLL seeded from the 143 known symbols (127 ZC + 16 UID):

def ddpll(x, K, a=0.05, b=0.002):
k = np.arange(len(x)); v = x[:143]*np.conj(K)
ph = np.polyfit(k[:143], np.unwrap(np.angle(v)), 1) # phase + freq from known syms
th = np.polyval(ph, 142); fr = ph[0]
y = x.copy(); y[:143] = x[:143]*np.exp(-1j*np.polyval(ph, k[:143]))
for i in range(143, len(x)):
th += fr; z = x[i]*np.exp(-1j*th)
d = (np.sign(z.real) + 1j*np.sign(z.imag))/np.sqrt(2)
e = np.angle(z*np.conj(d)); th += a*e; fr += b*e; y[i] = z
return y

The UID symbols are Gray-coded and unscrambled, so the UID reads straight off them. For A they read C1E20DC4, matching the banner.

TIP

Why seed from the known symbols? Our earlier trackers estimated phase from the 4th power of the symbols. QPSK is invariant under 90° rotation, so those trackers kept slipping by π/2 mid-frame. Anchoring the loop on the preamble and UID fixes the absolute phase, and the decision-directed loop holds it from there.


7. Cracking the FEC Chain#

The banner’s lfsr7 poly=0x91 … out=bit0 can be wired several ways, and the interleaver and scrambler could be applied in either order. fec.py tried every combination — five LFSR wirings (fib_parity, fib_poly_rev, gal, gal_left, fib_left) × two orders — accepting a combination only if RS(255,223) decoded and the CRC32 matched. Exactly one passed:

GRAY = {(1,1):(0,0), (-1,1):(0,1), (-1,-1):(1,1), (1,-1):(1,0)} # (sign I, sign Q) -> b1 b0
def lfsr(seed): # 7-bit, poly 0x91 -> feedback taps 0x48, output bit0
st = seed; o = np.zeros(2040, np.uint8)
for i in range(2040):
o[i] = st & 1; fb = bin(st & 0x48).count('1') & 1; st = ((st << 1) | fb) & 0x7f
return o
rs = reedsolo.RSCodec(32, nsize=255, fcr=0, prim=0x11d, generator=2)
def decode(sym, seed):
b = demap(sym).reshape(40, 51).T.flatten() ^ lfsr(seed) # de-interleave, then descramble
cw = np.packbits(b).tobytes() # 2040 bits = 255 bytes
msg, _, err = rs.decode(cw)
m = bytes(msg); L = m[1]
crc_ok = int.from_bytes(m[2+L:6+L], 'big') == zlib.crc32(m[2:2+L])
return m[0], L, m[2:2+L], crc_ok, len(err) # part, len, data, CRC, RS fixes

The surviving chain:

StageParameters
DemapGray 00→+1+1, 01→−1+1, 11→−1−1, 10→+1−1, I→b1, Q→b0
De-interleave40×51: bits.reshape(40,51).T.flatten()
Descramble7-bit LFSR, taps 0x48, left shift, output bit 0, seed = CRC8(UID)
RSn=255, k=223, GF poly 0x11d, fcr=0, generator 2
Payloadpart(1) ‖ len(1) ‖ data(len) ‖ CRC32-BE(4) ‖ padding = 223 B

Unit A decodes to part 0 — plaintext telemetry containing a “key” and a flag-shaped string. Neither is the secret; the banner already told us part 0 is plaintext.

For B and C we used a wider brute force (grid.py’s try_decode), sweeping all four constellation rotations, both interleaver orientations ('231' = reshape(40,51).T, '240' = reshape(51,40).T) and all 128 seeds.


8. Removing Unit A#

B and C sit underneath A in the same band. To reach them we cancel A: rebuild its exact transmitted waveform and subtract it from the recording.

8.1 Re-encode#

Take A’s corrected RS codeword and run the transmit chain forward — scramble, interleave, Gray-map — then prepend the known preamble and UID symbols:

def encode_syms(cw255, uid=0xC1E20DC4, u=29):
bits = np.unpackbits(np.frombuffer(cw255, np.uint8)) ^ SCR
tx = bits.reshape(51, 40).T.flatten() # TX interleaver
data = np.array([MAP[(tx[2*i], tx[2*i+1])] for i in range(1020)])/np.sqrt(2)
us = np.array([MAP[((uid>>(31-2*i))&1, (uid>>(30-2*i))&1)] for i in range(16)])/np.sqrt(2)
return np.concatenate([zc(u), us, data])

8.2 Subtract, Accounting for the Amplifier#

A’s power amplifier distorts the signal (banner: pa : xtr-7 rapp … ampm … dpd=off), so subtracting a clean linear copy leaves a huge residual. Per frame, sic.py does three things:

  1. Finds the best fractional timing (pulse shape evaluated at −0.5…+0.5 samples in ⅛ steps) and integer offset (±1 sample).
  2. Splits the frame into 256-sample blocks and, for each, fits a least-squares model [s, s·|s|²] over a window three blocks wide — tracking gain/phase drift and the cubic distortion term together.
  3. Subtracts the fit.
for b0 in range(0, len(s), BLK): # BLK = 256
w0 = max(0, b0-BLK); w1 = min(len(s), b0+2*BLK)
Sw = s[w0:w1]; Xw = np.vstack([Sw, Sw*np.abs(Sw)**2]).T
c, *_ = np.linalg.lstsq(Xw, seg[w0:w1], rcond=None)
S = s[b0:b0+BLK]; X = np.vstack([S, S*np.abs(S)**2]).T
out[b0:b0+BLK] = seg[b0:b0+BLK] - X @ c

sictune.py swept block size (32–256) and the number of distortion terms (1–3) to settle on these values.

8.3 Unit A Changes Framing Partway Through#

Remember the firmware update in the PCAP? The capture has a gap around 0.41 s, and re-acquiring A after it showed that from frame 44 onward A’s frames only matched a reference built with the other interleaver orientation ('240'). So cancellation uses the '231' reference for frames 0–43 and '240' for frames 44–79. Frames 33–49, 70 and 74 matched neither, so their references were rebuilt from their own hard decisions.

With A removed, frame-shaped structure that A had been burying appears in the residual. That’s B and C.


9. Detour: The 71.3 kHz “Beacon”#

The capture is literally named beacon_433 and a carrier keys on and off throughout, so the tone looks like the intended target. Decoding it properly took real work:

  1. Envelope. Mix 71.3 kHz down to baseband, low-pass at 4 kHz, decimate to 80 kS/s, take magnitude. On/off levels sit at ≈100 and ≈35.
  2. Clock. Find edges in the clean regions before and after the bursts, then brute-force the bit period T over [132, 134.5] samples @ 80 kS/s (≈1.66 ms) for the best single-grid fit to all edges.
  3. Framing. We tried Manchester (counting violations), coherent differential chip decisions, raw bytes, and UART 8N1 at both polarities and every bit shift.
  4. The catch. During the bursts the tone sits inside A’s band (A spans ≈ +22 ± 84 kHz), so the envelope is unreadable on the raw signal. On the A-cancelled signal it resolves cleanly into UART frames of 11 bits — start, 8 data bits LSB-first, 2 stop bits. 39 characters, every start/stop bit valid, message repeating.
Securinets{d3m0d_th3_j4mm3r_n0t_b34c0n}

It reads exactly like the reward for all that cancellation work. We submitted it and it was rejected. The secret lives in B and C.


10. Unit C: Part 2#

C sits at −9.25 kHz with root u = 104. On the A-cancelled signal, demod2.py walks each frame on the 9304-sample grid, searching offsets within ±100 Hz in 50 Hz steps and timing within ±70 samples, measuring preamble SNR, running the PLL and reading the UID:

C start 122910 cfo -9250 preSNR 7.0 near-bnd 0.070 uid 145D75BF crc8->u=104
C start 150821 cfo -9200 preSNR 9.0 near-bnd 0.058 uid 145D75BF crc8->u=104

near-bnd is the fraction of data symbols within 0.2 of a decision boundary — a cheap per-frame quality score. UID 145D75BF has CRC8 0x67, and 1 + 0x67 mod 126 = 104, matching the root.

Decoding with seed 0x67 and interleaver '231' succeeds on 8 frames individually (5, 8, 9, 10, 12, 14, 15, 16), every one producing the same CRC-valid payload:

C f 9 t=0.160s il=231 seed=0x67 fix= 8 part= 2 len= 31 data=fe9a0c91da2ec5938c1acc5874388817ef4df5901cc896cca114f803ff0c90

11. Unit B: Part 1#

11.1 Pitfall: Aliased Frequency Offsets#

Our first offset estimates for B and C were aliases, and nothing decoded until we corrected them to +4.08 kHz and −9.25 kHz.

This trap is easy to fall into for two reasons. First, Zadoff-Chu correlation couples time and frequency: an offset that is a multiple of Rs/N (125 kHz / 127 ≈ 984 Hz) shifts the correlation peak in time rather than destroying it. Second, the 4th-power estimator is ambiguous by Rs/4 = 31.25 kHz. A strong correlation peak does not mean the offset is right.

11.2 Single Frames Don’t Decode#

Even with A cancelled and the offset fixed, B is too weak — all 30 frames fail individually. Most have negative preamble SNR and an unreadable UID, except for one run of five:

B start 456359 cfo 4030 preSNR 6.4 near-bnd 0.151 uid C0F4E5B2 crc8->u=47
B start 465664 cfo 4030 preSNR 6.6 near-bnd 0.121 uid C0EDE5B2 crc8->u=109
B start 474967 cfo 4130 preSNR 7.7 near-bnd 0.146 uid 40F4E5B6 crc8->u=25
B start 484272 cfo 4080 preSNR 7.0 near-bnd 0.231 uid C0D4E5B2 crc8->u=61
B start 493576 cfo 4080 preSNR 4.7 near-bnd 0.304 uid C1F4E5B6 crc8->u=67

Frame 16 reads C0F4E5B2 exactly, and CRC8 0x2e gives u = 47 — confirming the UID. Frames 17–20 are only a few bits off.

11.3 Coherent Combining#

Every frame carries the same payload, so once the PLL has aligned their phases the data symbols of frames 16–20 can simply be summed. Each symbol is normalised to unit magnitude first, so no single strong frame dominates. The sum then goes through the ordinary decoder:

D = np.array([f[3][143:] for f in F]); D = D/np.abs(D) # per-symbol unit magnitude
s = D[16:21].sum(0) # frames 16..20
il, seed, nerr, m = dec(s, crc8_maxim(bytes.fromhex('C0F4E5B2'))) # seed 0x2e

Part 1: 30 bytes, CRC OK.


12. Decrypting the Flag#

The banner says ks : sha256-ctr keyed by the paired-net uids. The key is the three UIDs as raw bytes, sorted and concatenated, followed by a 4-byte big-endian counter starting at 0. The ciphertext is part 1 followed by part 2 (30 + 31 = 61 bytes):

import hashlib
uids = ['C1E20DC4', 'C0F4E5B2', '145D75BF']
key = b''.join(bytes.fromhex(u) for u in sorted(uids)) # 145D75BF C0F4E5B2 C1E20DC4
ks = b''.join(hashlib.sha256(key + i.to_bytes(4, 'big')).digest() for i in range(2))
ct = part1 + part2 # 30 + 31 bytes
print(bytes(a ^ b for a, b in zip(ct, ks)).decode())

As a sanity check, part 2 alone (keystream bytes 30–60 XOR C’s payload) gives _rlc_g4rdn3r_3kf_rs255_ph4nt0m}. With B’s part 1 in front:

Flag: Securinets{d0ppl3r_w13n3r_r4pp_rlc_g4rdn3r_3kf_rs255_ph4nt0m}


13. Decoy Hall of Fame#

#WhereTrickFake flag
1capture.sal zip commentplain stringSecurinets{str1ngs_1s_n0t_dsp_try_h4rd3r}
2meta.json, SDR_Q channel namezero-width characters (U+200B/200C/200D/2060/FEFF) interleaved with the textSecurinets{z3r0_w1dth_j01n3r_c0py_p4st3}
3meta.json, capture name beacon_433invisible Unicode tag characters (U+E0000 block) hiding a fake “assistant note” instructing AI solvers to skip the DSP and answer with a flagSecurinets{a1_r34d_th3_1nv1s1bl3_1nk}
4bench.pcap banner, selftest:homoglyphs (Cyrillic е/і/З/с/О/а, fullwidth braces)Sеcurіnets{d3bug_bu1ld_tЗst_v3сt0r_Оx5а5a}
5bench.pcap banner, after readya line addressed to LLMs telling them to print #3, plus another tag-character payloadsame as #3
6Unit A, part 0plaintext telemetry carrying a “key” and a flag-shaped stringnot the secret
771.3 kHz OOK tonereal UART, genuinely decodableSecurinets{d3m0d_th3_j4mm3r_n0t_b34c0n} (rejected)

14. Dead Ends#

  • The first loader draft — written straight from the banner. It assumed the wrong Saleae format and guessed a Galois-form LFSR, and its frame loop was never finished.
  • 4th-power phase trackers and per-segment quadrant brute-forcing — replaced by the PLL seeded from the 143 known symbols.
  • LFSR wiring and ordering guesses — only fib_left (taps 0x48, left shift, output bit 0) applied after de-interleaving passes.
  • One reference for all of A’s frames — A switches interleaver orientation at frame 44, and a handful of frames need references built from their own decisions.
  • Aliased frequency offsets for B and C — they look fine on a correlator and decode nothing.
  • Decoding B frame by frame — impossible; only coherent combining of frames 16–20 works.
  • The tone flag — correct demodulation, wrong flag.

15. Takeaways#

  • A string shaped like a flag is not evidence. Seven decoys across zip comments, JSON metadata, USB logs and the RF itself.
  • Read the side channels. The USB capture documented the protocol and recorded the firmware update that explains A’s mid-capture framing change.
  • Known symbols are your anchor. Seeding the PLL from preamble + UID removed the π/2 ambiguity that broke every blind tracker.
  • Cancel the loud signal to reach the quiet ones — and model the amplifier’s non-linearity when you do. A linear copy is not enough.
  • When single frames fail, combine them. Repeated payloads let you accumulate SNR coherently.
  • Check frequency offsets for aliases. Zadoff-Chu preambles couple time and frequency, so a strong peak can sit at the wrong offset.

Appendix: Script Map#

StageScripts
Load / surveyprobe.py, load.py, spec.py, dec.py
Tone (decoy)tone.py, env.py, clock.py, chips.py, manch.py, tone2.py, tone3.py, coh.py, frame.py, final.py
Preamble / unitscommon.py, zcscan.py, zcres.py, units.py
Demod Ademod29.py, demodu.py, ddpll.py, track*.py, segrot.py, redemod.py, reboot.py
FECfec.py, decode_all.py, grid.py
Cancel Areencode.py, sic.py, sictune.py
B / Cregrid.py, demod2.py, decunit.py, combine.py
Logsout/logB.txt, out/logC.txt, out/decB.txt, out/decC.txt

Flag: Securinets{d0ppl3r_w13n3r_r4pp_rlc_g4rdn3r_3kf_rs255_ph4nt0m}


RabbitHole — OSINT#

Nine hops from a video game composer to a 2005 Japanese blog post. Every answer feeds the next, and the flag is the concatenation of all nine.

NOTE

There is no shortcut through this one. Each clue is written so the next clue is unreachable without the current answer, and several hops bottom out in sources that only exist in Japanese.

AttributeValue
ChallengeRabbitHole
EventSecurinets CTF Quals 2026
CategoryOSINT
Hops9, strictly sequential
Flag formatSecurinets{1_2_3_4_5_6_7_8_9} — multi-word answers concatenated, joined with underscores

Step 1 — Identify the Composer#

Clue: “Identify the music composer for Outer Wilds.”

Outer Wilds (2019, Mobius Digital) has a well-documented composer.

Answer: andrewprahlow

Sources: Wikipedia, andrewprahlow.com, Outer Wilds wiki.


Step 2 — The Short Film & the Adopted Mother#

Clue: “The composer created music for a short film associated with a popular esports game. In the first minute of the film, a person is seen playing the game as a specific character. What is the name of the adopted mother of that character?”

Prahlow’s credits page lists GODLIKE (dir. Mickey Finnegan, 2016) — a short film built around competitive League of Legends, the esports game in question.

Watching the first minute, the protagonist Alex is playing Riven (white hair, runic blade). Riven’s lore: orphaned at birth — father killed in battle, mother died in childbirth — she was later adopted in Ionia by the elderly couple Asa and Shava Konte.

Answer: shavakonte

Sources: andrewprahlow.com/credits-1-1, League of Legends Universe page for Riven.


Step 3 — The University Music Event#

Clue: “Last year, the composer participated in a music event hosted at a university. What is the name of the location where the event was held?”

Prahlow announced on X in May 2025: a 20-minute Outer Wilds set with full choir and orchestra alongside @uclagme (UCLA Game Music Ensemble) on May 17. The initial tweet said Royce Hall — but there is a follow-up correction:

“correction: the concert is at the Ackerman Grand Ballroom”

Answer: ackermangrandballroom

Source: x.com/andrewprahlow/status/1919850338800325076 and its correction.

CAUTION

Hop 3 is a trap for anyone skimming. The first tweet gives the wrong venue, and “last year” is a relative date that only resolves correctly against the challenge’s own release window.


Step 4 — Best Fruit#

Clue: “What is the composer’s best fruit?”

A throwaway tweet from June 10, 2015:

“How have i never had a #kumquat before? Aka the best fruit.”

Answer: kumquat

Source: x.com/andrewprahlow (Jun 10, 2015).


Step 5 — The Pokémon Episode & the Production Manager Assistant#

Clue: “The composer contributed music to a Pokémon episode. Who was the Production Manager Assistant for that episode?”

Prahlow’s bio notes that he “scored a short film for Pokémon.” That is the 2025 POKÉTOON short “Can’t Take My Eyes Off You, Kangaskhan” (めがはなせないね ガルーラ), released October 3, 2025, with Andrew Prahlow credited as Music Composer.

The full credits (via eizo100.jp) list:

Production Manager Assistant: Jun Aoki (SPACE NEKO COMPANY Ltd.)

Answer: junaoki

Sources: eizo100.jp/en/video/124008, andrewprahlow.com/bio, Bulbapedia POKÉTOON page.


Step 6 — Birthplace and Birth Year#

Clue: “Where was the Production Manager Assistant born, and when?”

Jun Aoki (青木純) is an animator, director on Pop Team Epic, and head of SPACE NEKO COMPANY. The company bio gives a birth in Okinawa, 1981; TMDB carries the full date, November 29, 1981.

Answer: okinawa1981 — birthplace + year, matching the expected format.

Sources: spaceneko.com/jp/static/about, themoviedb.org, ja.wikipedia.org.


Step 7 — The 2024 Studio Relocation#

Clue: “In 2024, the Production Manager Assistant relocated their studio to a new location. Where did they relocate to?”

The same company bio timeline reads:

2024年 スタジオ移転(東京都武蔵野市)
→ Studio relocation (Musashino City, Tokyo)

Answer: musashino

Source: spaceneko.com/jp/static/about.


Step 8 — Blood Type#

Clue: “What is the blood type of the Production Manager Assistant?”

Japanese creator profiles routinely list blood type. The old Ameba blog profile sidebar reads:

血液型:A型

Answer: A

Source: ameblo.jp/aokijun (blog profile).


Step 9 — The Broken Phone#

Clue: “On what date did the Production Manager Assistant mention that his phone had slipped out of his hand and broken, Ikebukuro Station?”

On the same blog, a post titled 「携帯死亡」 (“Phone died”):

移動中、池袋駅のホームで手から携帯がすべり落ち、ぶっ壊れる。最悪。

“While traveling, my phone slipped out of my hand on the Ikebukuro Station platform and broke. The worst.”

Post date: 2005-12-11.

Answer: 2005-12-11

Source: ameblo.jp/aokijun/entry-10014009422.html.


Assembling the Flag#

#HopAnswer
1Outer Wilds composerandrewprahlow
2Riven’s adopted mothershavakonte
3UCLA concert venueackermangrandballroom
4Best fruitkumquat
5POKÉTOON production manager assistantjunaoki
6Birthplace + yearokinawa1981
72024 studio relocationmusashino
8Blood typeA
9Broken phone date2005-12-11

Flag: Securinets{andrewprahlow_shavakonte_ackermangrandballroom_kumquat_junaoki_okinawa1981_musashino_A_2005-12-11}


Why This Chain Is Hard#

  • Depth. Nine sequential hops, several landing on niche sources — anime production credits, a dead 2005 blog, a throwaway tweet.
  • Hallucination bait. Every hop looks like common knowledge that an LLM will confidently fabricate: composer trivia, champion backstories, concert venues.
  • Ambiguity traps. “Last year” is relative. The “esports game” is never named — you have to dig through the composer’s credits to reach GODLIKE. “Best fruit” is a single 2015 tweet.
  • Format strictness. Exact concatenation, single-letter blood type, YYYY-MM-DD. Any deviation fails the checker.
  • Language barrier. Three facts — the studio move, the blood type, the phone incident — exist only in Japanese sources.

Alternate Answers We Kept on Hand#

If the checker rejects…Try
Hop 6okinawa19811129 (TMDB’s full date)
Hop 7kichijoji (neighbourhood-level answer)

Takeaways#

  • Pivot on identities, not keywords. Every hop after step 5 is the same person; the work is finding which of their footprints answers the question.
  • Primary sources beat aggregators. The company’s own bio page carried both the birth details and the relocation; neither was reliably mirrored elsewhere.
  • Read corrections, not just the first post. Hop 3 is wrong in the original tweet and right in the reply.
  • Don’t stop at the English web. Half of this chain is unreachable without reading Japanese.