Day 3 - Web Exploitation - boolean-blind SQL injection in a search box - all three modes CORRECT on 2026-10-19 via the homestation Commander (same backlog pickup as Permission to Slip; agent-6 did not claim it on the daily release).
| Mode | GM result | Flag | Author of the solving method |
|---|---|---|---|
| Single Solve | CORRECT | a2a8205565ab4d70a2ff470f81544c84 | team shared pool (earned pre-this-run via a peer session; our own extraction confirmed the same proof_cb727cb1... value from the same target bytes) |
| Time Trial | CORRECT | 9331abb4b47fa1f59a38f73dc17617d4 | Commander (homestation) - parallel extractor, 5.6s end-to-end |
| Mastery | CORRECT | 0ade81cc896e517a8c259612a1b54d01 | Commander (homestation) - mixed SQLi + deploy-backup |
GM evidence for every mode: the real op=submit response had "result":"CORRECT" and solved_by.name:"the team solver". Nothing was guessed; every proof came from the live target's own data via a boolean-blind oracle.
Prompt and lane
Prompt (verbatim from the GM start response):
The search box trusts you a little too much. Ask the database one pointed question and read the row it was never supposed to show you.
Category: Web Exploitation. Same shared live-Web slot as the IDOR challenge (one live Web target per team at a time). GM delivers a live target URL in targets[0].url per started attempt. Mastery hands back TWO targets at once.
Why it looks SQLi but behaves oddly
The target's /lookup?q=... (Single, Mastery T2 at /search; Time Trial at /explore) runs a Python string-interpolated WHERE col LIKE '%' + q + '%' query on an SQLite backend. Three things trip up a naive sqli attempt:
--is accepted but every attempt to close the string with a matching quote leaves the trailing%'the server appends disabled. The classic' OR '1'='1'-- commentpayload returns zero rows: the server's%'is still there, so the query becomes... OR '1'='1'-- comment%'and the trailing fragment lives in a comment line that the SQL parser rejects as a syntax error (the whole statement errors; a Flasktry/except Exceptionswallows it and renders "Nothing matched.").' OR '1'='1(no closing quote) returns 3 rows, but so does' OR 'a'='b'and so does' OR 'anything'='x. That is becausepill LIKE '%' OR '<any>'='<any>%'isTRUE OR <stringcmp>; theLIKE '%'half always matches every row, so the OR fires regardless of the right-side comparison. It is NOT a real boolean oracle; it is a false witness, and we burned several minutes on it earlier.UNION SELECTwith any column count 1..15 returns 0 rows. The actual underlying SQL joins multipleOR col LIKE ...clauses with the sameqsubstituted into each, so a UNION in position 1 leaves the second and third LIKE patterns syntactically broken (trailing%'), and the whole statement errors. The entire UNION branch is a dead end on this app.
The real primitive: boolean sandwich via AND + LIKE prefix
The payload shape that is actually a boolean oracle is:
' AND (<predicate>) AND (1) LIKE '1
Server-interpolated: WHERE pill LIKE '%' AND (<predicate>) AND (1) LIKE '1%'. Parses cleanly because:
pill LIKE '%'is a trivial match (TRUE for every row).<predicate>is a free boolean. If TRUE, the AND chain survives; if FALSE, the row is filtered out.(1) LIKE '1'plus the server-appended%'becomes(1) LIKE '1%', a prefix LIKE comparing the integer1cast to text ('1') against the pattern1%. That is TRUE unconditionally and gives the trailing%'a safe landing.
TRUE path -> the three visible rows render on the page (Record NNN x3). FALSE path -> "Nothing matched." That is the oracle.
Pattern variants used in the live solve:
' AND (expr)<=<n> AND (1) LIKE '1for binary search over an integer.' AND UNICODE(SUBSTR((<subq>),<pos>,1))<=<code> AND (1) LIKE '1for a per-char binary search over printable ASCII (log2(95) == 7 requests per character).' AND (<subq>) LIKE '<prefix>as a prefix match: server-appended%'turns the prefix into<prefix>%'which is a LIKE prefix wildcard. (NOT used for blind-char extraction because_is a LIKE single-char wildcard and silently matches anything.)
Backend fingerprinting (SQLite)
sqlite_master lookups returned non-empty; pg_tables and information_schema. tables returned empty (both FALSE on the oracle). The schema was:
articles(id INTEGER, title TEXT, topic TEXT, summary TEXT, visibility TEXT)
4 rows total, 2 distinct visibility values ('public' and 'private'), one private row. That private row's summary holds proof_<32 hex>.
Extraction pipeline
Per target:
- Fetch
/with aUser-Agent: curl/8. The home page's<form action='...'>names the search endpoint for this variant (/lookup,/search, or/explore). All variants share the same backend. - Confirm the oracle:
' AND 1=1 AND (1) LIKE '1-> 3 rows.' AND 1=2 AND (1) LIKE '1-> 0 rows. - (OPTIONAL - skip once the pattern is proven across variants) extract the schema via
sqlite_masterandpragma_table_info('articles'). - Length of the private summary via binary search on
LENGTH((SELECT summary FROM articles WHERE visibility='private' LIMIT 1)). - Per character (parallel, ThreadPoolExecutor concurrency 24-32): binary search on
UNICODE(SUBSTR((<subq>),<pos>,1)). Characters are independent of each other (each is addressed by position), so the whole 38-character string extracts in one logical round. - Submit the recovered
proof_<hex32>to the GM under answer_idproof-1.
Measured end-to-end against the Time Trial live target: 5.6 seconds, 38 chars, ~270 HTTP requests. The target enforces a 60-second per-attempt window server-side (the challenge's declared "untimed" refers to Hall-of-Fame ranking, not target liveness - the target returns 410 Gone after about 60s of inactivity from the first fetch). The parallel extractor fits comfortably.
Mastery: SQLi + deploy-backup mixed
Mastery returns two targets at once:
- Target 2 is the exact SQLi template; same schema, same
articlestable, same private-row extraction (3.8s end to end on 2026-10-19). - Target 1 is a non-SQLi "Tiny Uptime" status page. Same deploy-backup leak as Permission to Slip Mastery T1:
robots.txtdisallows/archives-a7ba9/, which is a directory index servingrelease-settings-92415c.json.bak. That JSON'ssupport_notefield is the proof.
Submit both proofs in one op=submit.
Common dead-ends and traps
' UNION SELECT ... --on this app: never works regardless of column count or comment style. Skip the UNION branch after one confirmation.' OR '1'='1' ...as a boolean oracle: FALSE witness because the leading LIKE '%' is always TRUE. Use AND, not OR, to build an oracle on this app._and%are LIKE wildcards: the extractor must useGLOBranges orUNICODE(SUBSTR(...))numeric compare per character, not aLIKE 'prefix+ch'prefix match (early build was fooled into returning____for every position).- The Time Trial attempt reconnects to the SAME attempt_id across
op=startandop=retrywhile the target isactive. Once the target expires (410 Gonefrom its own route prefix,elapsed_ms>60000), re-issuingop=startafter the GM clears the attempt yields a fresh target with a new URL and route. - The homestation
ctf_gm.GM().start(challenge="/time_trial")refuses by default (TimedStartRefused). SetHUNTRESS_TIMED_START=commander-overridewhen the Commander owns the solve (the Web lane has no rehearsed prebrief).
Lessons learned (banked)
- Boolean-blind SQLi on an interpolated
LIKE '%'||q||'%'sink has one universal payload shape' AND (<predicate>) AND (1) LIKE '1. The sandwich absorbs the server-appended%'into a safe LIKE prefix. See [[kb-web-blind-sqli-sandwich]]. - Boolean oracle only works with AND, not OR, when the leading LIKE pattern is
'%'(always TRUE). An OR branch would short-circuit TRUE regardless of the right-hand predicate. - Parallel per-character extraction is essential for timed targets. 60s is tight for sequential blind-SQLi; a ThreadPool with 24-32 workers lands the proof in ~5s.
- Mixed-mode Mastery is a real possibility on the Web track: one SQLi target
- one deploy-backup target under one attempt. Run the attack ladder on each target independently.
- Robots.txt deploy-backup pattern recurs - exactly the same
Disallow:-> directory index ->*.json.bak->support_noteon both days 02 and 03 Mastery T1. Promote it to a first-class web primitive. - Project's existing web_contracts.py doesn't cover live-API families - its four executors (hash/jwt/config/http) are all offline over static artifacts. The live-Web lane needs its own primitives file (
web_live_primitives.py- committed this pass), hooked into theweb_solvers.LADDERfor next time.
Files
- (this writeup)
Query_Me_This_Writeup.md - GM record (authoritative):
challenges/Day 03/_gm_record/web-exploitation-day-03-sqli.homestation.jsonl - Working probes (purged after solve per
feedback-purge-gm-downloads): home-page snapshots per variant and lookup/search/explore response samples were written into the project directory, used, and removed. The small discovery scripts (sqli_fast_parallel.py,probe_endpoint.py,sqli_master.py) are retained one level up as reference implementations of the primitive now generalised inchallenges/_lib/web_live_primitives.py.