CVE-2025-14847 - MongoDB Unauthenticated Memory Leak Exploit

A proof-of-concept exploit for the MongoDB zlib decompression vulnerability that allows unauthenticated attackers to leak sensitive server memory.

A flaw in MongoDB's zlib message decompression returns the allocated buffer size instead of the actual decompressed data length. This allows attackers to read uninitialized memory by:

1. Sending a compressed message with an inflated uncompressedSize claim
2. MongoDB allocates a large buffer based on the attacker's claim
3. zlib decompresses actual data into the start of the buffer
4. The bug causes MongoDB to treat the entire buffer as valid data
5. BSON parsing reads "field names" from uninitialized memory until null bytes

# Basic scan (offsets 20-1MB)
python3 mongobleed.py --host <target>

# Exact-size dump (single probe)
python3 mongobleed.py --host <target> --dump 10MB

# Dump with a small window around the target (more reliable)
python3 mongobleed.py --host <target> --dump 10MB --dump-window 2048

# Auto-tune dump size/window for best yield
python3 mongobleed.py --host <target> --auto

# Loop until stopped (Ctrl+C)
python3 mongobleed.py --host <target> --loop

# Optimized scanning strategy (sampling + hot offsets)
python3 mongobleed.py --host <target> --optimize --loop --decode

# Interactive TUI hexdump viewer
python3 mongobleed.py --host <target> --tui --loop

# Deep scan for more data
python3 mongobleed.py --host <target> --max-offset 50000

# Custom range
python3 mongobleed.py --host <target> --min-offset 100 --max-offset 20000

# Decode URL/unicode/base64 previews
python3 mongobleed.py --host <target> --decode

When you run with just --loop --decode, these defaults apply:

[*] mongobleed - CVE-2025-14847 MongoDB Memory Leak
[*] Author: Joe Desimone - x.com/dez_
[*] Target: localhost:27017
[*] Scanning offsets 20-50000
[*] Output: leaked_localhost_27017_20250101_120000.bin

[+] offset=  117 len=  39: ssions^\\x01\\xf4r\\x9a\\x2aYDr\\xc3\\x90
[+] offset=16582 len=1552: MemAvailable:    8554792 kB\nBuffers: ...
[+] offset=18731 len=3908: Recv SyncookiesFailed EmbryonicRsts ...

[*] Total leaked: 8748 bytes
[*] Unique fragments: 42
[*] Saved to: leaked_localhost_27017_20250101_120000.bin

• To request ~20MB buffers, target that size with --max-offset 20971520 (and --buffer-extra 0).
• To keep a smaller scan range but request larger buffers, use --buffer-extra so doc_len + buffer_extra equals your target size.
• Use --dump for a single probe at an exact claimed size (e.g., --dump 10MB).
• --dump uses an automatic window when --dump-window 0; override it for tighter or wider scans.
• --auto ignores manual --dump/offset ranges and picks a size/window based on quick probes.
• --auto-legacy uses a legacy scan to bias auto toward hot offsets.
• In --auto mode, the window auto-expands after several empty passes to keep finding new leaks.
• Use --auto-mode size if you want maximum leak size instead of speed.
• --optimize reuses connections per worker and rotates between sampling and dense scans.
• Use the hit= token printed alongside a leak with --hit to replay the same probe.
• Add --hit-wiggle to probe around the hit offset for more leaks.
• Use --hit-backoff to slow down hit replay loops.
• The effective cap is the server’s maxMessageSizeBytes; values above it will be rejected before parsing.

• Controls
• ↑/↓ move one row
• PgUp/PgDn jump one screen
• n next hit
• p previous hit
• o toggle search mode (linear vs optimized)
• v toggle preview pane
• d toggle decode in preview
• space pause/resume refresh
• q quit
• Behavior
• Auto-sizes rows and bytes-per-row to terminal width/height
• Preview pane shows the longest decoded fragment from the current refresh cycle
• Next/Previous hit search scans forward/backward until a leak is found, then jumps the view (toggle mode with o)
• Requires Textual: pip install textual (auto-resize supported).
• Workflow:
  1. Run --auto-mode speed --loop --decode --optimize to find hot offsets and capture a hit= token.
  2. Launch TUI with --hit <token> --tui to browse a live view around that offset.

Compared to the original release, this fork adds:

• Multi-threaded scanning with configurable worker count
• Continuous scanning modes (--loop and --max-empty-passes)
• Automatic, timestamped output naming with host/port
• Unique-only output writing with append support across runs
• ASCII-safe previews to avoid mangled console output
• Optional decoding of URL/unicode/base64 previews (--decode)
• Escaped character cleanup for decoded previews (\\uXXXX, \\xNN, \\n, etc.)
• Hexdump fallback for non-readable leaks and --hex override
• Automatic dump size/window tuning (--auto)
• Optimized scanning mode with sampling + hot offsets (--optimize)
• Connection reuse per worker in optimize mode
• Hit replay tokens with --hit and wiggle/backoff support
• Textual-powered TUI live memory browser (--tui) with MC-style layout and refresh
• Rich-formatted console logging
• Additional tunables (--buffer-extra, --timeout, --preview-bytes)
• KB/MB size parsing for offsets and buffer size />

A Docker Compose file is included to spin up a vulnerable MongoDB instance:

docker-compose up -d
python3 mongobleed.py

The exploit crafts BSON documents with inflated length fields. When the server parses these documents, it reads field names from uninitialized memory until it hits a null byte. Each probe at a different offset can leak different memory regions.

Leaked data may include:

• MongoDB internal logs and state
• WiredTiger storage engine configuration
• System /proc data (meminfo, network stats)
• Docker container paths
• Connection UUIDs and client IPs

• OX Security Advisory
• MongoDB Fix Commit

• Hamid Kashfi (@hkashfi)
• Codex (OpenAI)
• Joe Desimone - x.com/dez_

This tool is for authorized security testing only. Unauthorized access to computer systems is illegal.

• Python 3
• Rich (pip install rich)