Day 17 - Fetch Quest (script-analysis-day-17-download-graph)
Lane: script-analysis | Agent: agent-1 | Machine: homestation | All three modes CORRECT on first contact.
GM flags (accepted, not redacted)
| Mode | Flag | Time | Ranked |
|---|---|---|---|
| Single Solve | Tokence800023e48a64210072bbb90303a5dc | 3.4 s | no (untimed) |
| Time Trial | Tokenfeb912ad28260a66478467d163be7356 | 4.9 s | yes |
| Mastery | Tokendc2407bc65a1ecacd452a08321811c3f | 13.0 s | yes |
Solved under handle the team solver. No 409/429/503, no retry, no probe loop, no HELD residue - AUTOSOLVE opened each timed window only after the readiness gate passed on the previous mode's live result.
The challenge
GM prompt:
The source code says where it would fetch from, and the generated network evidence says where the requests actually went. Reconcile both into one download graph, and find the indicator that shows up on only one side of the story.
Artifact: artifact_0.bin (challenge artifact, not published), a 114808-byte AES-encrypted ZIP (password infected), 127 inert script members - 22 .sh, 14 .py, 12 .js, 10 .vbs, 9 .php, 8 .mjs, 5 each of .cmd/.lua/.pl/.rb, 4 each of .ps1/.bat/.mof/.ts/.au3/.applescript/.proj/.ahk.
Field counts: Single and Time Trial 1139 fields each; Mastery 5695 fields (5 independent sets of 1139, set1. ... set5. prefixes, each set its own re-randomized member zip).
The "network evidence" framing is flavour: there is no pcap or log member. Every fixture is self-describing - the descriptor inside each file is both the source-side intent and the recorded-side result, so "reconciling both sides" is exactly decoding the Material_<id>/Codec_<id> pair and reading the steps/stage_profile/url keys out of the recovered descriptor. The "indicator on only one side" is the per-fixture URL/campaign/mutex/wallet value that appears only in the decoded descriptor and never in the plaintext of the carrier.
How it was solved
No new reversing was needed. The banked lane spine (challenges/_lib/autosolve_script.py, driven by challenges/_lib/autosolve.py) handled the whole thing:
ctf_gm.unzipopens the AES zip with passwordinfected; Mastery's outer zip is split into its five per-set member zips byautosolve_script._set_blobs.- Each member is decoded via the Day 7 structured extractor, falling back to the generic
Material_/Codec_path (generic_decode) and then tosolve_poly_d8.extra_answersfor the PHP / host-containerv=3descriptors and the advanced-bash fixtures. 52 members fall through the primary extractor (AppleScript, AutoIt, AutoHotKey, Lua, Perl, Ruby, HTA/WSF/SCT hosts) and are answered by the generic/poly paths - they are "undecodable" only to the first extractor, not to the solver. field_formats.confirm_allbanked 44 new tokens on the Single and 5695 on Mastery, so each later mode started with today's formats already confirmed.
Offline replay reproduces the GM result exactly: 1139/1139 resolved, 0 unresolved (scripts/replay_single_offline.py -> scripts/single_answers.json).
Today's new layer
The novelty keys were a new polyglot carrier family: AutoHotKey (.ahk), 44 fields across four fixtures (benign-utility, downloader, launcher, staged-chain). It is a carrier, not new knowledge - the descriptor keys are unchanged, so the generic path answered all 44 with no code change. This confirms the standing hypothesis (now n=6) that a self-describing new carrier costs nothing while the descriptor schema holds.
Recovered AutoHotKey values (Single):
| id | value |
|---|---|
polyglot-auto_hotkey-benign-utility-steps | acquire_remote>dispatch_process |
polyglot-auto_hotkey-benign-utility-attack | none |
polyglot-auto_hotkey-downloader-steps | decode_embedded>write_file>dispatch_process |
polyglot-auto_hotkey-downloader-attack | T1105 |
polyglot-auto_hotkey-downloader-url | https://auto_hotkey-6381587e8d.example.invalid/stage |
polyglot-auto_hotkey-launcher-steps | profile_host>decode_configuration>dispatch_language |
polyglot-auto_hotkey-launcher-attack | T1059 |
polyglot-auto_hotkey-staged-chain-steps | decode_configuration>return_configuration |
polyglot-auto_hotkey-staged-chain-role | staged_chain |
All four report host-profile=standalone and a next-language handoff (powershell / jscript / batch / vbscript), which is what makes the cross-language edges of the "download graph".
The predicted candidate scheduled-task xml was already prepared and also appeared; nothing in the candidate list was missing.
The three headline PowerShell samples
| id | value |
|---|---|
resource-scenario | resource_hijacking |
resource-attack | T1496 |
resource-url | https://cdn-0597d83ddd.example.invalid/1d0a52188fa1.bin |
resource-wallet | wallet_18f77064a52f4328c4e396cdbe06952c |
resource-mutex | Global\HuntressLab_545caad2f0a4 |
collection-scenario | synthetic_collection |
collection-attack | T1555.003 |
collection-url | https://cdn-0dfc6f5541.example.invalid/950800736965.bin |
collection-file | update-0dfc6f5541.dat |
collection-user-agent | Mozilla/5.0 HuntressLab/3.2 |
impact-scenario | synthetic_impact |
impact-attack | T1486 |
impact-url | https://cdn-2c72ba33b8.example.invalid/7198cd0eb43e.bin |
impact-extension | .locked-2c72b |
impact-note | RECOVERY-2c72ba33b8.txt |
collection-attack is the one hand-steered answer: the GM's own question text overrides the usual "Data from Local System" instinct and tells you to treat the browser credential store as the central behavior, so the answer is T1555.003 (Credentials from Web Browsers), not T1005. Follow the GM's wording, not the family default.
Bash-advanced fixtures
16 advanced-bash specimens, every one role=advanced_syntax, attack=T1059.004, with the feature and primitive read straight out of the fixture's own material (self-describing, so a new feature is free): ansi_c_quoting, array_slicing, associative_arrays, brace_expansion, dynamic_file_descriptor, getopts_state, here_string, heredoc_reconstruction, indirect_expansion, nameref_dispatch, process_substitution, recursive_walk, regex_conditional, subshell_pipeline, token_fragmentation, trap_handler.
Distinct ATT&CK ids across all 1139 fields: T1041, T1053, T1053.003, T1059, T1059.003, T1059.004, T1059.005, T1059.006, T1059.007, T1105, T1140, T1486, T1496, T1555.003, plus the literal none for benign utilities.
Takeaways
- Flavour text can describe two sides of a story that are physically one artifact. Do not go looking for a pcap that the digest's member list does not contain; the descriptor carries both halves.
- A new polyglot carrier (AutoHotKey) with the unchanged descriptor schema needs zero code. Recognize, reuse, do not rebuild.
- When a GM question explicitly redirects the technique mapping ("treat X as the central behavior rather than Y"), that sentence is the specification.
collection-attack=T1555.003.
Files
scripts/summarize_day17.py- read-only field/family census of task + both contracts.scripts/replay_single_offline.py- offline (no GM) replay proving 1139/1139.scripts/single_answers.json- the full accepted Single answer set.