Day 21 - Actions Speak Louder (script-analysis)
- Challenge id:
script-analysis-day-21-github-actions-persistence - GM name:
Actions Speak Louder - Lane: script-analysis
- Agent: agent-1 (solver
the team solver) - Result: ALL THREE MODES CORRECT at first contact, no retries, no spent variants.
Flags (GM-accepted)
| Mode | Flag | Time to CORRECT | Ranked |
|---|---|---|---|
| Single Solve | Tokenf752390395b3ba2e6e3c5e58464d1f37 | 4.2 s | n/a (untimed) |
| Time Trial | Token58dbeb1e5074989890cbfadcfa8b7450 | 4.4 s | yes (hall_of_fame_eligible) |
| Mastery | Token3f3d1012ac7b029a796fbc86fb039672 | 9.4 s | yes (hall_of_fame_eligible) |
GM receipts: challenges/Day 21/_gm_record/script-analysis-day-21-github-actions-persistence.homestation.jsonl (13 records: 3 starts, 3 artifact pulls, 3 submits, 4 free status reads). Every submit returned HTTP 200 result: CORRECT on the first batch - no partial-fields, no incorrect-answer, no 409/429.
Challenge prompt (GM's own words)
A GitHub Actions workflow decides who can trigger it and which runner inherits the trust. Identify each trigger and trust boundary, follow the static handoff into the next language, and find the job that keeps coming back.
Shape of the attempt
| Mode | artifact bytes | sha256 (short) | answer_ids |
|---|---|---|---|
| single_solve | 133,401 | e005f70c2a553e72 | 1,373 |
| time_trial | 133,499 | 5ec55f1f6322ecb8 | 1,373 |
| mastery | 667,173 | 9e7d732431331805 | 6,865 |
Artifact is the usual evidence.zip (challenge artifact, not published) of inert, self-describing script fixtures - 148 distinct members referenced by the Single's question set. Timed deadlines were server-provided (Time Trial 2026-10-21T13:01:38.598Z, Mastery 2026-10-21T13:03:03.728Z), both routes reported unique_attempts_remaining: 31; we spent zero of them because the first start was also the winning attempt.
Question families in the Single (1,373 ids):
polyglot 660 bash 168 host 132 python 126
jscript 72 batch 72 vbscript 72 nodejs 56
resource/collection/impact 5 each
Per-fixture field set is the familiar 7-11 field block: role, attack, adapter, campaign, url / file-path, stage-count, steps, host-profile, obfuscation, stage-profile, next-language.
Today's new layer: GitHub Actions workflow YAML as a carrier
The novelty words from the pre-claim digest were actions, github, louder, speak and the lane's NEXT-LAYER candidate list had no match - the layer was correctly flagged UNPREDICTED. It turned out to be a new carrier, not a new general-purpose language:
- 44 new ids under the prefix
polyglot-github_actions-* - artifact members
polyglot-github_actions-<role>.ymlinsideevidence.zip(challenge artifact, not published) - roles present:
benign-utility,downloader,dropper,launcher,uploader,persistence-descriptor,staged-chain(the standard role wheel, now wearing a workflow file)
This is exactly the D8-D10 pattern (PHP polyglot -> stager graph -> WMI MOF) continued: a CI/CD persistence descriptor is just another host container around the same Material_<id> / Codec_<id> self-decoding descriptor. The prompt's "who can trigger it and which runner inherits the trust" maps onto the fixture's host-profile / stage-profile fields; "the job that keeps coming back" is the persistence_descriptor role; "follow the static handoff into the next language" is next-language.
Representative accepted answers for the new family:
polyglot-github_actions-benign-utility-role = benign_utility
polyglot-github_actions-benign-utility-attack = none
polyglot-github_actions-benign-utility-adapter = local_parser
polyglot-github_actions-benign-utility-campaign = GITHUB_ACTIONS-95b5fe73
polyglot-github_actions-benign-utility-file-path = /tmp/huntress-github_actions/.95b5fe73fc977d09.stage
polyglot-github_actions-benign-utility-stage-count = 2
polyglot-github_actions-benign-utility-steps = acquire_remote>dispatch_process
polyglot-github_actions-benign-utility-host-profile = standalone
polyglot-github_actions-benign-utility-obfuscation = base64
polyglot-github_actions-benign-utility-stage-profile = two_stage_download_execute
polyglot-github_actions-benign-utility-next-language = powershell
polyglot-github_actions-downloader-attack = T1105
polyglot-github_actions-downloader-adapter = http_client
polyglot-github_actions-downloader-url = https://github_actions-11083c4649.example.invalid/stage
polyglot-github_actions-downloader-steps = decode_embedded>write_file>dispatch_process
polyglot-github_actions-downloader-obfuscation = hex_base64
polyglot-github_actions-downloader-stage-profile = three_stage_decode_write_launch
polyglot-github_actions-downloader-next-language = jscript
polyglot-github_actions-launcher-attack = T1059
polyglot-github_actions-launcher-adapter = process_builder
polyglot-github_actions-launcher-steps = profile_host>decode_configuration>dispatch_language
polyglot-github_actions-launcher-obfuscation = fragmented_hex
polyglot-github_actions-launcher-stage-profile = cross_language_handoff
polyglot-github_actions-launcher-next-language = batch
polyglot-github_actions-staged-chain-adapter = operation_graph
polyglot-github_actions-staged-chain-steps = decode_configuration>return_configuration
polyglot-github_actions-staged-chain-obfuscation = plain_hex
polyglot-github_actions-staged-chain-stage-profile = single_descriptor
polyglot-github_actions-staged-chain-next-language = vbscript
All indicators are inert by construction (*.example.invalid, /tmp/huntress-*). Nothing was executed - static descriptor extraction only.
Today's two new bash-advanced features
As predicted by the standing hypothesis ("bash-advanced features are self-describing; a new one costs nothing while the descriptor keys are unchanged"), Day 21 added two:
bash-advanced-ansi-c-quoting-feature = ansi_c_quoting
bash-advanced-ansi-c-quoting-primitive = ansi_c_escape_string
bash-advanced-ansi-c-quoting-role = advanced_syntax
bash-advanced-ansi-c-quoting-attack = T1059.004
bash-advanced-ansi-c-quoting-steps = recognize_syntax>reconstruct_intent
bash-advanced-ansi-c-quoting-campaign = BASHX-A3FEC887
bash-advanced-array-slicing-feature = array_slicing
bash-advanced-array-slicing-primitive = indexed_array_slice
bash-advanced-array-slicing-role = advanced_syntax
bash-advanced-array-slicing-attack = T1059.004
bash-advanced-array-slicing-steps = recognize_syntax>reconstruct_intent
bash-advanced-array-slicing-campaign = BASHX-ED142F0B
The pre-brief gap report had listed all 100 bash-advanced ids (50 Time Trial + 50 Mastery) as "no extractor". That report scores against the named extractor table, not the generic path - the generic Material_/Codec_ reader answered every one of them. The gap was a false positive, and no build was needed.
Method (CONTRACT per mode)
RECOGNIZE -> the claim digest fingerprinted the artifact as the lane's known Material_<id>/Codec_<id> self-decoding fixture family (TWIST constant since Day 6).
BIND -> the started attempt's own evidence.zip (challenge artifact, not published) bytes, per mode. Each timed mode derived every answer from its own mutated artifact; nothing was carried over from the Single except the method.
EXTRACT -> autosolve_script.py's language-agnostic descriptor path: _generic_material -> _layered_decode / _raw_decode -> generic_decode -> solve_fields_ext. This validates a printable v=1 descriptor and prefers the longest competing extractor, so the YAML workflow carrier needed no new LANGS table entry - confirming that from Day 8 onward a new carrier is free while the descriptor contract holds. field_formats.confirm_all banked 44 new value tokens before the Single and 220 more before Mastery.
CORRELATE -> role / ATT&CK precedence applied from the decoded ordered behavior graph (terminal step wins), as the attack questions spell out; next-language read straight from the handoff descriptor.
VALIDATE -> every field resolved locally with zero unresolved ids before any GM call. Evidence rating: SURE - the values come from the decoded descriptor bytes, not from inference, and the GM graded all three full batches CORRECT on the first submit.
SUBMIT -> one batched submit per mode (1,373 / 1,373 / 6,865 answers), each well under the 4 MiB limit. Order was Single -> Time Trial -> Mastery, with Mastery started only after Time Trial came back CORRECT.
Timing / policy notes
- No
run_for_timewas ever sent (standing user directive). Both timed routes were still gradedhall_of_fame_eligible: truebecause this route supportsrun_for_time, so a plain first start is the ranked attempt. - Both timed starts used the EARLY path, justified by the readiness gate: the prior mode was already CORRECT live on this same challenge, i.e. the same compute was proven against real bytes - no rehearsal wait and no idle window.
- No HELD residue, no probe loops, no
HUNTRESS_TT_BISECT, no 429/409/503 anywhere in the record. unique_attempts_remainingstayed at 31 for both timed routes: not a single variant was burned.- Downloaded artifact bytes were purged after the solve, per the own-lane rule; the GM record and this writeup are the retained evidence.
Lessons to carry forward
- Day 21 = carrier, not language - fifth consecutive confirmation of the D8+ hypothesis. GitHub Actions workflow YAML joins PHP polyglot, stager graph and WMI MOF as a host container around the unchanged descriptor. The lane's carrier universe should now list
.yml/.yamlCI/CD workflow (GitHub Actions, T1053-adjacent CI persistence) as SEEN and handled by the generic path. - The pre-brief "no extractor" gap report over-reports. It checks named extractors only. When the missing ids are plain
Material_/Codec_fixtures (here: 100 bash-advanced ids), the generic reader already covers them - verify against the generic path before building anything. - Prompt-to-field mapping is reliable. "Who can trigger it / which runner inherits trust" -> host-profile and stage-profile; "the job that keeps coming back" -> persistence_descriptor role; "static handoff into the next language" -> next-language. Reading the GM's prose first saved a schema hunt.
- The readiness gate works as designed. A live-CORRECT prior mode on the same challenge is the strongest possible proof that the same compute will finish inside the window; taking the EARLY path on that basis produced two ranked 4-9 s solves with zero variants spent.