🌐 Fail Open Friday — Day 24 Web Exploitation
- Challenge ID:
web-exploitation-day-24-fail-open-friday - Lane:
web-exploitation - Station: Homestation (HS Web lane)
- Solved by: HS (All three modes CORRECT per GM)
- Writeup salvage source: HS-held notes captured 2026-10-28 after D28 recovery; preserved into the repo so WS + HS both have the method on record.
Results
| Mode | Result | Game Master flag |
|---|---|---|
| Single Solve | CORRECT | Tokena689349485af38f2d0811e8d1ce25977 |
| Time Trial | CORRECT | Tokenb3284507e5bd961148bc944dd9239bc7 |
| Mastery | CORRECT | Token5ee6d6324289fde6a04ce17b5020bfef |
All attempts were untimed. Time Trial and Mastery both started with run_for_time=false and no deadline.
Challenge prompt
"The remote policy decoder is part of an access decision. A valid public decision is documented in the application; investigate what happens when that decision cannot be decoded."
There were no downloadable artifacts — each attempt assigned live target(s). SS and TT asked for one proof from a single target; MM asked for proof-1 through proof-5 from five independent targets.
Method summary
Single Solve + Time Trial (same shape)
- GET
/api/clueson the assigned target. The response identified the feature route and documented the valid public decision shape:remote:<id>:<base64url-JSON>with a documentedstandarddecision. - Sending a MALFORMED value with the documented prefix (e.g.
remote:40b12d94f983:not-base64!) to the feature route returned HTTP 200 witheffect: "exceptional_condition_fail_open"and aproof_<hex>value. - Submit that
proof_<hex>under the exactanswer_idreturned by GM start (proof-1in both SS and TT).
Mastery (5 targets, each a different vuln class)
| Answer | Target technique | Proof |
|---|---|---|
proof-1 | robots.txt → disallowed /snapshots-89561/ → directory listing exposed *.bak JSON | Proofproof_89f15f7eba154759b1cc003a967d44ea |
proof-2 | Home-page case_session cookie → /api/timeline reveals case 13636 → /api/cases/13636 returns the proof | Proofproof_8c9d93bc9dd4475b825c933027ea36aa |
proof-3 | Search-param SQL injection ') OR '1'='1'-- → hidden note with proof | Proofproof_2427ca8bdce6461f9d63a24bedf36a94 |
proof-4 | Reader ?doc= traversal → /run/huntress/objective.txt → proof | Proofproof_d6657ad4fea147c295e2c8c7edc2ea80 |
proof-5 | /api/clues → documented remote:6fe40b6a2a03: prefix + malformed base64 → proof | Proofproof_4db29cadca7e4de5a4ede92ab346818c |
Lessons / faster-next-time
- Each mode's targets are per-attempt. The route suffixes, session cookies, export IDs, and proof values change between assignments. Always use the current attempt's assigned URLs.
- Follow visible clues first.
/api/clues,robots.txt, linked index pages, and the public challenge prompt beat guessing paths. - Preserve session cookies. APIs that look standalone (
/api/timeline,/api/cases/<id>) may require a home-page session cookie first. - For Mastery, inspect each target independently. Mastery can be a MIXED bundle of vuln types (archive leak, IDOR, SQLi, LFI, audit-event fail-open) — do not assume one exploit applies to all.
- Submit the exact
answer_idfrom the start response (proof-1…proof-5here, exact names only). - Don't guess answers; derive from live responses. The GM is never a value-probe.
Script
scripts/huntress_fail_open_probes.py — reconstructed probes for SS, TT, and the five Mastery target types. Read-only: prints proof values; does NOT submit to GM. The agent submits exact proofs under exact answer IDs via autosolve.py.
HS lane adoption
HS web lane (web-exploitation) handles this via agent-6/7/8 with CID-hash ownership. The same autosolve_web.py adapter can learn these 5 probe shapes (archive leak / IDOR / SQLi / LFI / audit fail-open) once; TT/MM mutate per attempt but shape is identical.