Tactical LogHuntress CTF 2026 // team archive

Days / Day 25 / Binary RE / Out of Bounds

DAY 25//LANE: BINARY RE//STATUS: CORRECT x3

Out of Bounds

Binary RE SSTTMM Modes reported CORRECT in this write-up

Day 25 - Out of Bounds (binary-reverse-engineering)

  • Challenge id: binary-reverse-engineering-day-25-synthetic-vulnerability-root-cause
  • Lane: binary-reverse-engineering
  • Agent: agent-2 (solver identity on the GM: Blandman3)
  • Status: ALL THREE MODES CORRECT (GM-accepted)

GM-accepted flags

ModeResultFlagRanked / HoF
Single SolveCORRECTToken13dcd4ad0a3c3169600e37502146233bn/a (untimed)
Time TrialCORRECTToken586f897fa780caeab6d38a11454289d3ranked, hall_of_fame_eligible: true
MasteryCORRECTToken66c2fe196b585321142a5ed72cd03422ranked, hall_of_fame_eligible: true

Challenge statement

Prompt (GM):

Somewhere in here, the code reaches where it shouldn't. Prove the bounds error is real, trace its reachability, and compare against the corrected twin that learned its limits.

Artifact: evidence.zip (challenge artifact, not published), ~35.7 MB per attempt, delivered per started attempt (Single 1 zip, Time Trial 1 zip, Mastery 9 zips at ~35.7-35.8 MB each). statement_sha1 for Single: e8a4fbabeeeeb347fba694dd5e6cad759ddb6808.

Shape: the cumulative Day 01-25 inventory board. 300 answer ids in Single and Time Trial - 12 slots per day prefix d01_* .. d25_*; Mastery is the same 300 slots across 9 sets (set1.d01_* .. set9.d25_*) = 2700 answers under one timer.

Today's new layer: a vulnerable/corrected ROLE PAIR in Java

The d25 members are a two-member JAR pair distinguished only by a basename role suffix:

  • the unique d25-- member whose basename ends vulnerable.jar (challenge artifact, not published)
  • the unique d25-- member whose basename ends corrected.jar (challenge artifact, not published)

The title theme ("Out of Bounds", synthetic bounds error vs. the corrected twin) is narrative framing for the pair; none of the 12 graded d25 slots asked for a CWE, a line number, or an exploitability verdict. The slots were the lane's standing inventory shapes re-pointed at the pair:

d25 slotWhat it askedAccepted answer
d25_artifact_sha256_1SHA-256 of member 1 in ZIP central-directory ordercc14d96c8bd784dfcaae87b602ea5bc576d213ad3b8a4c70bc26f2de83982553
d25_artifact_sha256_2SHA-256 of member 2 in ZIP central-directory order8fab8d498c6695c8605689d13e72e2d798e0a242c214d0313146cf429a0c5eac
d25_analysis_vulnerable_formatcontainer class of the vulnerable.jar (challenge artifact, not published) memberjar
d25_analysis_corrected_formatcontainer class of the corrected.jar (challenge artifact, not published) memberjar
d25_analysis_config_campaignHCFG campaign, required byte-equal across BOTH role membersBR25-000
d25_analysis_config_domainHCFG domain, both role membersnode-000.day25.example.invalid
d25_analysis_config_mutexHCFG mutex, both role membersGlobal\HuntressBR25_000
d25_analysis_config_urlHCFG url, both role membershttps://node-000.day25.example.invalid/api/BR25-000
d25_analysis_corrected_config_campaignHCFG campaign, corrected.jar (challenge artifact, not published) onlyBR25-000
d25_analysis_corrected_config_domainHCFG domain, corrected.jar (challenge artifact, not published) onlynode-000.day25.example.invalid
d25_analysis_corrected_config_mutexHCFG mutex, corrected.jar (challenge artifact, not published) onlyGlobal\HuntressBR25_000
d25_analysis_corrected_config_urlHCFG url, corrected.jar (challenge artifact, not published) onlyhttps://node-000.day25.example.invalid/api/BR25-000

The genuinely new structural wrinkle is the cross-member equality precondition on the four d25_analysis_config_* slots. The GM's own question text spelled it out:

Decode both d25-- role members and require their config objects to be byte-for-byte equal before using the common value.

So the vulnerable and the corrected twin carry the same HCFG config; the pair exists to be compared, and the four "shared" slots are only answerable once both decodes agree. The four corrected_config_* slots are the same four keys sourced from the corrected member alone - which is why both groups returned identical values.

HCFG extraction path (unchanged spine, Java container)

Per the GM's guidance, for a JAR:

  1. Parse every classfile constant pool in the JAR.
  2. For Base64-alphabet UTF-8 constants of at least 80 characters, Base64-decode.
  3. Accept the first CRC32-valid HCFG record.

HCFG record layout (constant across all 25 days):

0..3   ASCII "HCFG"
4      version
5      one-byte XOR key
6..7   little-endian payload length
8..11  little-endian CRC32 of the DECODED payload
12..   payload

XOR every payload byte with the key, verify CRC32, parse the decoded UTF-8 as JSON, read the top-level property. This is the same record the lane has decoded since Day 01; only the carrier changed (Day 06 Java constant pool, Day 16 wasm data section, Day 25 Java constant pool again, in a role pair).

Why this needed zero new code

challenges/_lib/autosolve_binary.py (2355 lines) contains no vulnerable / corrected / role-pair logic and did not need any. Two already-banked generic mechanisms covered the day:

  1. _Q_MEMBER_PATS member resolver (autosolve_binary.py:1398-1406) - built on Day 21 for "Use the unique d21-- member whose basename ends plugin.cmxs." It honours a basename-tail token only when the token contains a dot and exactly one member's basename ends with it. vulnerable.jar (challenge artifact, not published) and corrected.jar (challenge artifact, not published) are exactly that shape, so each role slot bound to the right member with no new case.
  2. The by-name fact route - a slot whose field name matches a binary_facts fact name is answered SURE from the parsed facts (container class, SHA-256 by index, HCFG keys).

That is the whole reason the day went clean on first touch: today's "new layer" was a new naming convention over already-modelled structures, not a new structure. Recognize, then reuse.

Execution timeline (UTC, 2026-10-25)

Release at 13:00:02.049Z. All three modes CORRECT within 68 s of release, every answer derived from that attempt's own bytes.

EventTimeNote
Single start13:00:02.8840.83 s after release; untimed, no deadline
Single submit CORRECT13:00:11.296300 answers in one submit; 8.4 s end to end
Time Trial start (RANKED)13:00:11.831deadline 13:01:32.115Z -> 80.3 s window; unique_attempts_remaining 63
Time Trial submit CORRECT13:00:19.330300 answers; 7.5 s used, 72.8 s slack; attempts remaining still 63
Mastery start (RANKED)13:00:19.671deadline 13:10:15.241Z -> 595.6 s window; unique_attempts_remaining 15
Mastery submit CORRECT13:01:10.1932700 answers, 9 evidence zips (~322 MB) downloaded and parsed; 50.5 s used, 545.0 s slack; attempts remaining still 15

Readiness gate on both timed modes: the prior mode was already CORRECT live on this challenge, so the same compute was proven on real bytes and no rehearsal wait was needed. No run_for_time request was sent (standing policy); the plain first start was the ranked attempt and both came back hall_of_fame_eligible: true.

Notably the ranked first-touch 409 (timed_intent_not_available) that hit this lane on Days 09-12 did not recur: both timed starts were accepted ranked on the first call, and neither submit consumed a unique attempt (63 -> 63, 15 -> 15).

Evidence rating

SURE for all 12 d25 slots and all 288 carried-over slots: every value was computed from the downloaded bytes of the attempt being graded (ZIP central-directory order for member indexing, magic + parsed container for the format tokens, CRC32-validated HCFG for the config keys), and the GM accepted all three modes outright with a flag and no remaining list - no partial-field grading, no probes, no guesses.

Operational notes

  • GM calls made: 3 starts, 11 artifact downloads, 3 submits. No retries, no practice, no polling, no 429/409/503.
  • Own lane only; no other lane's challenge was started, probed, or downloaded.
  • Nothing was executed from the artifacts - static parsing only (ZIP central directory, classfile constant pools, HCFG decode).
  • No artifact copy was retained on disk for this lane (autosolve streamed each evidence.zip (challenge artifact, not published)); the accepted SHA-256 values above are the durable evidence record.

Findings, confirmed hypotheses, and what could be better

Finding 1 (pipeline defect, affects every lane): outcome.json records ERROR:no_outcome on a clean sweep

All four Day 25 lanes finished with every mode CORRECT, yet all four _board/*.outcome.json files record "outcome": "ERROR:no_outcome", "detail": null:

binary-reverse-engineering-day-25-...     outcome= ERROR:no_outcome
metadata-forensics-day-25-hive-mind       outcome= ERROR:no_outcome
mobile-android-day-25-integrity-signals   outcome= ERROR:no_outcome
warmups-day-25                            outcome= ERROR:no_outcome

The claim digest showed the matching symptom in-line: AUTOSOLVE ALL: autosolve printed no outcome line then AUTOSOLVE OUTCOME: ERROR:no_outcome, immediately after three AUTOSOLVE <mode>: CORRECT lines and an AUTOSOLVE NEXT: every mode of this challenge is CORRECT line. So the per-mode results were correct and only the summary line was missing or unparsed.

Why this matters beyond cosmetics: the lane FAILURE HISTORY block in the claim digest is built from these outcome records. The binary lane already boots with error-other [OPEN] seen Day 12,17 (2x) - Unclassified adapter error: read the traceback tail. A perfect day that self-reports ERROR:no_outcome will keep inflating that counter and send tomorrow's agent chasing a traceback that does not exist. The authoritative success record is _board/<id>.autosolve.json (per-mode result + flag) plus done.json; the outcome file disagreed with both. Recommend fixing the outcome-line emit/parse in the autosolve-to-claim_next handoff, and until then treating ERROR:no_outcome as "unknown", not "error", when building the failure history.

Finding 2: the ranked first-touch 409 is no longer reproducing

The lane's timed-start-409 [FIXED 2026-10-10] entry (6 occurrences across Days 09-12) did not recur. Both timed starts were accepted ranked on the first plain call, came back hall_of_fame_eligible: true, and neither submit consumed a unique attempt (TT 63 -> 63, Mastery 15 -> 15). The fix is holding on a Day-25-era route.

Finding 3: timed windows are server-provided and much wider than the rehearsed assumption

Observed windows were not ~30 s. Time Trial gave 80.3 s (submitted with 72.8 s slack) and Mastery gave 595.6 s (submitted with 545.0 s slack, including nine ~35.7 MB downloads). The lane's timed-slot heuristic is tuned to TT 18s; that bound is conservative by roughly 4x on this route. Keep reading the deadline from each start response rather than planning against a fixed 30 s, exactly as the standing rule says.

Confirmed hypotheses

  • Cumulative evidence bundle: CONFIRMED again. evidence.zip (challenge artifact, not published) carries d01-- .. d25-- members and has grown to ~35.7 MB; Single and the timed variants share identical slot shapes (300 ids = 25 day prefixes x 12).
  • "Zero or one new slot per day": CONFIRMED in the stronger form. Day 25 added a new day prefix (12 slots) but zero new slot shapes - every d25 slot is an existing shape (artifact_sha256_N, analysis_<role>_format, analysis_config_<key>) re-pointed at a new member layout. The novelty budget went into member naming, not into question vocabulary.
  • Mastery mutation: CONFIRMED. 9 independent sets, 9 distinct evidence zips (distinct sizes and hashes), one timer, 2700 answers.

What could be better for this lane's category

  1. The candidate list missed today, and the miss was predictable. The digest said "no predicted candidate matches today's ids/artifacts/fields - treat today's layer as UNPREDICTED". Every prepared candidate was about new content (unasked HCFG keys, record geometry, PE/ELF header fields, new toolchains, malware id). Day 25 instead varied member selection: a two-member role pair inside one day prefix. Recommend adding a member-layout axis to NEXT-LAYER CANDIDATES - role/twin pairs (vulnerable vs corrected, before vs after, packed vs unpacked, debug vs release), N-member sets, and cross-member preconditions - since Day 15 (symbolized+stripped pair), Day 16 (direct+compact wasm pair) and Day 25 (vulnerable+corrected JAR pair) are now three instances of the same axis while the lane still models novelty as a content question.
  2. Cross-member preconditions deserve a first-class handler. Today's shared-config slots said "require their config objects to be byte-for-byte equal before using the common value". It happened to pass, so a solver that ignored the precondition and read either member still scored. A future variant that makes the twins differ would silently reward the wrong member. An explicit equality check (and a Hold when it fails) would convert a latent wrong-answer risk into a visible residue entry.
  3. The theme/question gap is worth keeping prominent. "Out of Bounds" plus "Prove the bounds error is real, trace its reachability" strongly implies vulnerability analysis; the graded slots asked for hashes, container tokens and config strings. Reading the narrative prompt as the task would have burned the untimed window decompiling Java for a CWE nobody asked about. The standing instruction to follow the GM's own per-question text over the title/prompt narrative paid for itself here.
  4. No artifact is retained for this lane. Autosolve streams evidence.zip (challenge artifact, not published) and nothing lands on disk, so post-hoc verification depends entirely on the GM record plus the accepted hashes. Keeping one copy of the Single's zip (untimed, already downloaded) would make a disputed field re-checkable without a refetch.