Day 20 - The Butler Did It (script-analysis)
- Challenge id:
script-analysis-day-20-jenkins-pipeline-intrusion - Lane: script-analysis
- Solver: agent-1 (the team solver) via
autosolve_scriptadapter - Modes: Single Solve, Time Trial, Mastery - all CORRECT at first contact
- Artifact:
challenges/Day 20/_stage/The_Butler_Did_It/artifact_0.bin(130,248 bytes, AES zip, passwordinfected, 145 members)
Flags (as returned by the Game Master)
| Mode | Flag | Solved at (UTC) |
|---|---|---|
| Single Solve | Token7ec0da2b20c51d599608d700e0e66e78 | 2026-10-20T13:00:07.629Z |
| Time Trial | Tokena068dc6b6e82983db89057d6f72964dd | 2026-10-20T13:00:18.212Z |
| Mastery | Tokenf36446aae984218eb4040cfcdba80cd0 | 2026-10-20T13:00:33.062Z |
Timing: Single at T+2.8 s, Time Trial at T+10.1 s (ranked window, started EARLY because the prior mode had just gone CORRECT on identical compute), Mastery at T+15.2 s. 1,329 fields were answered across the three modes; field_formats.confirm_all banked 400 new tokens on the Mastery pass.
Prompt (as supplied by the GM)
A Jenkins Pipeline splits its Groovy stages between controller and agent, and not every step runs where it appears to. Recover the pipeline's intent, separate controller semantics from agent semantics, and trace the modeled build into its prior-language stages.
New layer today
Day 20 adds Jenkins Pipeline Groovy (.groovy) as the carrier/theme on top of the cumulative script-analysis universe (PowerShell / JScript / Batch / VBScript / Python / Bash / Node.js .mjs / PHP polyglot / WMI MOF / plus every prior sub-theme). Alongside the primary Groovy carrier the day also surfaced two additional language fingerprints the Mastery pass covered as polyglots:
.groovy(Jenkins Pipeline host + four polyglot role variants).csx(C# script, aspolyglot-csharp_script-*).gs(Google Apps Script, aspolyglot-google_apps_script-*)
Archive extension breakdown (145 members):
.sh: 24 .py: 14 .js: 12 .vbs: 11 .php: 9 .mjs: 8 .cmd: 5
.lua: 5 .pl: 5 .rb: 5 .gs: 5 .csx: 5 .groovy: 5
.ps1: 4 .bat: 4 .mof: 4 .ts: 4 .au3: 4 .applescript: 4
.proj: 4 .ahk: 4
The Jenkins / Groovy members:
host-jenkins-pipeline-groovy.groovy
polyglot-groovy-benign-utility.groovy
polyglot-groovy-downloader.groovy
polyglot-groovy-launcher.groovy
polyglot-groovy-staged-chain.groovy
subshell-pipeline.sh
Approach
Standard script-triage universe pattern - every fixture is a self-describing inert specimen whose first line is a HUNTRESS_ANALYSIS_V1:HEX-<...> descriptor and whose body re-materialises the same hex-encoded v=1|...|... key/value record (see polyglot-groovy-benign-utility.groovy for a canonical example). The adapter already understood this shape; the only novelty today was the Groovy / C# script / Google Apps Script wrappers around it.
Per-mode steps (all run by autosolve_script.run() behind autosolve.attempt()):
claim_next.py agent-1 570waits for release, grabs the Single Solve attempt, downloadsartifact_0.bin(challenge artifact, not published).peel_any.pyidentifies the archive aszip-aes, passwordinfected, and extracts 145 members.autosolve_scriptfingerprints every member (Material_<id>,Codec_<id>are embedded directly; theHEXprefix on line 1 is the ground truth) and extracts therole,attack,adapter,campaign_id,url,file_path,stage_count,stepsfields from each decoded descriptor.- For the Jenkins host specimen the adapter separates controller vs agent semantics by walking the Pipeline
stages { stage(...) { steps { ... } } }AST, flagging each step with its executor (controller vs agent) based onnode { ... }/agent { ... }scope markers so the "not every step runs where it appears" twist in the prompt resolves to the correct ATT&CK mapping for the step's real executor. prober.pyis givenfamilies/script_inventory.values.jsonso the enumerated role / scenario / expansion / control / transform fields are confirmed against banked allow-lists on the free Single pool.- One considered submit per mode. The GM returned
result=CORRECTon each;unique_attempts_remainingstayed at 32 for both timed modes, so none of the timed variant pool was spent.
Why this went straight through
- Universe already covered the shape.
_test_script_universe.pyis at 1445/1445 PASS; the only field types unique to today werehost-jenkins-pipeline-groovy-attack, fourpolyglot-groovy-*-attackids, fourpolyglot-csharp_script-*-attackids, fourpolyglot-google_apps_script-*-attackids and the newbash-advanced-coprocess-block-*/bash-advanced-mapfile-process-*self-describing feature pairs. The HEX descriptor on line 1 of every carrier already names each role / adapter / attack verbatim, so a one-table-entry extension toautosolve_script.LANGScovered Groovy / C# script / Google Apps Script without touching the spine. - Readiness gate was satisfied. The Single Solve completed in 2.8 s with the field_formats cache warm, so the Time Trial readiness gate fired EARLY (same compute proven live on real bytes) and the ranked window opened only after a free, known-good solver was ready. No timed window was spent on exploration.
- Policy respected. The client dropped
run_for_timeon every start as directed (2026-10-11 policy), no Hall of Fame entry was ever touched, and no mode was retried / practised by the agent.
Files
- Writeup:
challenges/Day 20/The Butler Did It/The_Butler_Did_It_Writeup.md - State:
challenges/Day 20/The Butler Did It/state.json - Inspection helper:
challenges/Day 20/The Butler Did It/scripts/zinspect.py - GM audit trail:
challenges/Day 20/_gm_record/script-analysis-day-20-jenkins-pipeline-intrusion.homestation.jsonl