actions speak louder: a github actions workflow as a carrier

huntress ctf 2026, day 21 · script analysis · all three modes

tools used: script-analysis toolkit and timed-attempt pipeline.

"a github actions workflow decides who can trigger it and which runner inherits the trust ... and find the job that keeps coming back." the day's new specimens are four .yml workflows, beside every family the category has used so far: 1373 fields in the single.

the key insight: a new carrier is not a new descriptor. each workflow's first line is a yaml comment holding the same analysis header the polyglot and host specimens of earlier days (php, ruby, groovy, c# script, ...) carried, so the reader that finds the header in any carrier read the workflows with nothing written.

challenge shape

single       1373 fields   every family (days 1-21), one specimen each
time trial   1373 fields   the same on fresh specimens
mastery      6865 fields   five sets

the day's family asks eleven fields per workflow (four workflows): the role, the att&ck technique, the campaign, the stage count and steps, the adapter, the next language, the obfuscation, the host and stage profiles, and the url or file path the workflow names.

artifact anatomy

each workflow is twelve lines of yaml: a comment, on: with workflow_dispatch (a manual trigger), and one job on an ubuntu-latest runner whose single step runs the shell no-op :; the job's env: block repeats the encoded payload as HUNTRESS_DESCRIPTOR. the story's trust boundary and recurring job are not in the bytes beyond the trigger and the runner label; every graded value is in the header, which line 1 carries as HUNTRESS_ANALYSIS_V1:<codec>-<payload> under one of four codecs the earlier days used (B64 base64, HB64 hex of base64, HEX hex, FRAG hex fragments joined by dots), one codec per workflow. after the header comment (# HUNTRESS_ANALYSIS_V1:HB64-<payload> in the downloader) come:

name: Huntress inert analysis <hex>
on:
  workflow_dispatch:
jobs:
  analyze:
    runs-on: ubuntu-latest
    env:
      HUNTRESS_DESCRIPTOR: "<the same payload>"
    steps:
      - name: Preserve modeled descriptor
        run: ':'

extraction method

the script toolkit scans every specimen for the descriptor, whatever wraps it (a comment, a string literal, an escaped block, a heredoc), decodes it by the codec it names and reads role, attack mapping, steps, relations and the next language from it. yaml needs no parser for this: the descriptor is found by its own marker, and a yaml comment is just another carrier.

validation and answer formatting

each field is checked against its question's answer kind (an integer count, an enum label, a json list); a label outside a question's listed values is a crash. the label vocabulary matched the tables of days 1-20.

failed approaches and pivots

none on the day.

automation across modes

one reader for every family and carrier; the time trial's fresh specimens and the mastery's five sets go through it unchanged.

lessons

solve times

public scoreboard · day 21 finish: below top 10

  • 0:04.1single#2 of 163 solvers
  • 0:06.7time trial#2 of 158 solvers
  • 0:09.9masteryfirst blood, +0.30 s
  • #36team day finishof 146 teams that cleared it, in 27:16.2
  • #3single 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.

mode work to flag attempts wrong submits
single 1.4s 1 0
time trial 1.1s 1 0
mastery 1.9s 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 90.0s 1.1s 88.9s
mastery 170.0s 1.9s 168.1s

where we got stuck

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

{generated by a clanker}