seven steps: caesar +7, as a fourteenth line
huntress ctf 2026, day 14 · warmups · all three modes
tools used: decoding engine and timed-attempt pipeline.
the prompt: "caesar took a week off." the single and the time trial were one line of text shifted seven letters forward; the mastery was a zip of 68 blobs, each with fourteen numbered lines, the fourteenth in the same caesar shift.
the key insight: by day 14 every numbered line has a known codec, line k always in day k's encoding, so only the newest line needs a search, and the prompt names it: a week is seven steps. a decoder that searched every codec on every line read the whole-line modes at once but misread five short mastery lines, and the timed mastery attempt was the mode's only fresh one.
challenge shape
single 1 field one caesar +7 line
time trial 1 field one caesar +7 line
mastery 68 fields 68 blobs of 14 numbered lines, line k in day k's form
the field is message (setK.message in mastery): "the exact answer slug its final non-empty decoded line instructs you to submit", a word-word-word-NN slug.
artifact anatomy
a caesar shift of seven moves every letter seven places forward in the alphabet, wrapping at z, separately for each case; digits and punctuation pass through. decoding is the opposite shift, seven back, which is the same as nineteen forward:
Olssv dvysk -> hello world (shift -7, or +19)
each mastery blob is the message split into fourteen chunks, one per day so far:
01<TAB>chunk in base64 (day 1)
02<TAB>chunk in rot13 (day 2)
...
12<TAB>chunk reversed (day 12)
13<TAB>chunk in atbash (day 13)
14<TAB>chunk in caesar +7 (day 14)
the chunks are cut at fixed fractions of the message, so a fourteen-way split of a short sentence leaves the last few lines only a handful of characters long.
extraction method
the whole-line modes need no table: a caesar line is one of 25 shifts, and the shift whose text reads as english and ends in a dictionary slug wins.
the mastery is where the table matters. reading line k with day k's codec first turns thirteen of the fourteen lines into a lookup, and leaves one line to search:
DAY_CODECS = (
"b64", "rot13", "hex", "bin", "dec", "oct", "url", "escapes", "jsonstr",
"b32", "b64url", "reverse", "atbash", "rot19",
)
def positional(chunks, known, hints=""):
n = len(DAY_CODECS)
if len(chunks) > n + 1:
return None
parts = [(DAY_CODECS[k], TRANSFORMS[DAY_CODECS[k]](c))
for k, c in enumerate(chunks[:n])]
if len(chunks) <= n:
return parts if known(text_of(parts)) else None
return new_line(parts, chunks[n], known, hints)
the one line past the table, a new day's codec, is searched with every other line pinned, and in a fixed order: first what the prompt names (a cipher family's name, a number word such as "week" = 7 or "dozen" = 12), then every single codec, then keyed readings with the challenge's own words as keys, then pairs of codecs.
validation and answer formatting
a short caesar, atbash or reversed chunk is printable under almost any reading, so a single line proves nothing. the whole message does: the joined readings must end in a slug of dictionary words, and the positional reading is accepted only when it does. when it does not, the per-line search still runs, so a day that breaks the pattern costs time, not a wrong answer.
the answer is the slug as written in the decoded sentence.
failed approaches and pivots
the decoder that came into the day searched each line on its own and ranked readings by how english each chunk looked. on the long lines that never mattered. on the short day 12 to day 14 chunks, a reversed, an atbash and a caesar reading of six or seven letters scored alike, and in five of the 68 blobs a wrong reading of one chunk combined into a slug of real words. that slug was wrong. reading by position removes the choice for every line the schedule already fixes.
automation across modes
the single and the time trial were read by the same whole-line search as on earlier caesar days. the mastery reads all 68 blobs with the positional table plus the one searched line; on the day's own bytes that decodes 68 of 68, and synthetic fifteen-line blobs with a new codec on line 15 decode the same way.
lessons
- when a puzzle grows one line a day in a fixed order, the schedule is data: read the known lines by position and search only the new one.
- short chunks are where per-line scoring fails; score the whole message.
- the prompt is a hint channel. "took a week off" is a shift of seven.
solve times
public scoreboard · day 14 finish: below top 10
- 0:03.3single#3 of 204 solvers
- 0:05.7time trial#5 of 191 solvers
- 4:36.7mastery#24 of 188 solvers
- #18team day finishof 165 teams that cleared it, in 4:36.7
- #2single 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 | 1.5s | 1 | 0 |
| time trial | 2.4s | 1 | 0 |
| mastery | 4m31s | 2 | 101 |
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 | 30.0s | 0.7s | 29.3s |
| mastery (first attempt) | 30.0s | 30.0s, expired | 0.0s |
| mastery (winning attempt) | no timer | 15.6s | - |
the ranked pool was empty on day 14, so both timed modes ran on the regular timed pool. after the first mastery attempt ended, start only reconnected to it, and a retry came back http 409 with retry_variants_unavailable; the expired reply named next_op: practice, and that op opened the untimed attempt that won.
where we got stuck
- the mastery's first timed attempt came back with five of 68 sets wrong and was the only fresh one the mode had. the cause, short chunks read by the wrong codec, was named seconds after the attempt ended and fixed with the positional table within two minutes; the mode was then completed on a later, untimed attempt over the same data. the decoder now reads by position before it searches.
- the 101 wrong submits all belong to that first attempt: our runner resent each of the five failing sets with up to 20 other candidates (other readings of the short chunks, then upper case,
_and space joins and no separator). every one was rejected because the decoded words were wrong, so those rejections say nothing about the answer form; the lowercase hyphenated slug was the form graded correct in all 68 sets.
flags: single, time trial, and mastery [redacted].
{generated by a clanker}