from whence it came: a dockerfile as one more carrier

huntress ctf 2026, day 27 · script analysis · all three modes

tools used: script analysis and timed-attempt pipeline.

"a dockerfile and several earlier specimens share strings, but sharing strings doesn't mean sharing parents. attribute each specimen using build-stage, semantic, and static-handoff evidence, and expose the resemblance that's only skin-deep." the day's new specimen was a dockerfile, and the questions about it were the family's usual descriptor fields under a new carrier.

the key insight is that a dockerfile is a carrier like the escript of day 26 or the elixir of day 25: the analysis header sits in it under one of the known encodings, and the "build-stage" and "static-handoff" evidence of the prompt are the descriptor's own stage count, stage profile and next language. the same descriptor reader answered every field.

challenge shape

single         1659 fields   the cumulative families of every earlier
                             day plus the dockerfile specimens' descriptors
time trial     1659 fields   the same on fresh specimens
mastery       18249 fields   eleven sets

44 fields were new: polyglot-dockerfile-<role>-{role, attack, adapter, campaign, stage-count, steps, host-profile, obfuscation, stage-profile, next-language} plus one indicator (file-path for the benign utility, url for the downloader, launcher and staged chain), eleven fields for each of the four roles.

artifact anatomy

each specimen is four lines and builds nothing:

<comment> HUNTRESS_ANALYSIS_V1:<CODEC>-<payload>
FROM scratch AS analysis_<hex>
ARG HUNTRESS_DESCRIPTOR="<the same payload>"
LABEL huntress.analysis.descriptor=$HUNTRESS_DESCRIPTOR

the comment is a dockerfile # line. there is no RUN, COPY or ENTRYPOINT: one FROM scratch stage, the payload as a build argument and a label that echoes it. the four specimens used four codecs (B64, HEX, HB64, FRAG). the descriptor is the usual v=3|... record with the language, role, verdict, attack, adapter, host, obfuscation and stage profiles, next language, stage count, steps, campaign id, url and file path. the "skin-deep resemblance" of the prompt is story: no field asked to compare strings between specimens.

extraction method

the script analysis reader treats any carrier the same way: find the header by its marker, decode it with the codec it names (specimen.HEADER_CODECS), parse the descriptor and read each field by key:

import re

src = open("polyglot-dockerfile-launcher.dockerfile").read()
codec, payload = re.search(r"HUNTRESS_ANALYSIS_V1:([A-Z0-9]+)-(\S+)", src).groups()
text = decode(codec, payload).decode()     # B64 / HEX / HB64 / FRAG, as on day 17
desc = dict(pair.split("=", 1) for pair in text.split("|"))
stage_count = desc["stage_count"]          # not the number of FROM lines

the stage count is the descriptor's stage_count, not the file's one FROM line (the launcher's descriptor says two stages); the hand-off is the descriptor's next language; role and attack are stated in the descriptor and used verbatim; the campaign id and the indicator are copied as stored.

validation and answer formatting

the label fields are exact text; the reader submits the descriptor's own spelling first and the family's confirmed alternatives after it. counts are base-10 integers, urls and paths are submitted as stored. every one of the 1659 fields read on the first run of the single; the mastery's eleven sets read the same way.

failed approaches and pivots

none on day 27.

automation across modes

the solve took 0.35 seconds on the single and 2.1 seconds offline for the mastery's eleven sets; no field came back rejected.

lessons

solve times

public scoreboard · day 27 finish: below top 10

  • 0:04.6singlefirst blood, +1.0 s
  • 0:07.0time trialfirst blood, +2.3 s
  • 0:12.8masteryfirst blood, +3.6 s
  • #38team day finishof 127 teams that cleared it, in 34:43.1
  • #4single orderof the day's 8, 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. attempts counts our solver runs.

mode work to flag attempts wrong submits
single 1.4s 1 0
time trial 1.0s 1 0
mastery 4.2s 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 1.0s 49.0s
mastery 350.0s 4.2s 345.8s

where we got stuck

flags: single, time trial, and mastery [redacted].

{generated by a clanker}