Longer writing on the craft — detection engineering, reverse engineering, incident narratives, and the parts of this work that survive the NDA.
EST. MMXVIII2 PUBLISHED3 SEALED
READ 00% — FROM THE TOP
Before official results came out, anyone with curl could read 10.4 million candidate marks with zero login. The story, the flaw, and the fix.
One CSS proxy URL returned 3,376 bytes. Behind it: full database access to 20,600 student accounts, 22 AWS buckets, Slack admin tokens, and universal session forgery.
2026.10.2312 MIN READDISCLOSURECRAFT
Before official results came out, anyone with curl could read 10.4 million candidate marks with zero login. The story, the flaw, and the fix.
Two requests. No session, no token, nothing. The first returns 4,884 bytes. The second returns 24,193 bytes. The only thing that changes between them is four characters — 1=1 swapped for 1=2. Behind that difference sat a PostgreSQL database with 10,433,626 candidate records for India's biggest national examinations.
Candidates look up results by choosing an exam code and year; the browser sends parameters like examname=CISF2 and examyear=2025. Secure code uses prepared statements — user input handed to the database inside a sealed envelope, treated as text, never as orders. This backend did raw string concatenation instead: whatever arrived from the internet was pasted straight into the SQL. No prepared statement. Not one. No allow-list either. A single quote mark breaks out of the envelope, and everything after it runs as live instructions.
RESPONSE SIZE IS THE ONLY SIGNAL — NO ERROR TEXT, NO REFLECTION. ONE REQUEST PER BIT OF DATA.
A blind injection, playing 20 questions: anchor to a valid exam code and any yes/no question appended runs against the live production database. examyear had its own tells — IS NULL false meant the column exists; examname IS NULL crashed with a 500, out of scope. One handler, two unparameterized queries, two tables, one door.
First question worth asking: who is the application running as? The answer came back TRUE — postgres. Not a restricted web account. The PostgreSQL superuser: it can read any table in every database on the machine, read arbitrary files off the filesystem, execute operating-system commands. The public lookup page for civil service marks was connected with the master keys to the entire house. current_database() resolved to marksheet_system; the public schema held 73 tables.
Names, roll numbers, dates of birth, categories, paper-by-paper marks, interview scores. How do you pull text from a server that only says yes or no? Binary search — guess the character in seven questions, exactly like finding a number between 1 and 100. The proof was sixty lines of clean Python from the standard library. No sqlmap, no specialized suites. To confirm the vulnerability without taking unnecessary data, exactly one record was extracted — a real candidate whose details matched public merit listings. One. Nothing more.
The first clue was subtle: an extra quote mark crashed the page mid-banner at exactly 691 bytes, right after the logo rendered. A crash proves the door exists; it doesn't hand you keys. The breakthrough was the anchor — locked onto CISF2, a tautology and a contradiction produced reliable, distinct byte counts every single time, and the crash became a high-speed channel.
Fingerprinting took three tests: 1::text='1' evaluated TRUE (PostgreSQL cast syntax); @@version returned a 500 (MySQL syntax); 'a'||'b' concatenated cleanly. PostgreSQL confirmed, no verbose errors needed. ORDER BY 6 lived, ORDER BY 7 died — six columns — and a six-NULL UNION SELECT drew the auto-logout redirect: UNION-capable too. And the moment 10.4 million records plus superuser rights are confirmed on a public, unauthenticated portal, the research phase is finished. You do not scrape. You do not test writes. You do not dump.
On 23 October I compiled the dossier — reproduction steps, payloads, remediation advice — and submitted it to CERT-In under reference CERTIn-56442226. Six days later the UPSC engineering team deployed the fix. Legacy portals at /marksheet/csm_q_marks_2020/ and /marksheet/csp_premark_2020/ shared the architecture and were fixed too. The remedy required no magic: parameterized queries — prepare() and bindParam() — and never regex filtering or blacklists. Input must remain data, never instructions.
On the evening of 29 October I tested the oracle once more. 1=1 and 1=2 now return the identical byte count. The differential has collapsed. The door is closed.
Ten million records. Forty columns. One unquoted parameter. Six days from disclosure to fix.
"Trust is the expensive part. Everything else is scripts."
वसुधैव कुटुम्बकम् — the world is one family.