thc-BloodMiami

AI chat-box pentesting tool, written in C. Fires language (SQL/NoSQL/XSS/SSTI) and prompt injection payloads at a chat interface and flags the ones that bite — DB error strings, reflected script markers, model jailbreak signals. Fast, threaded, libcurl based.

What it does

• Language injection (-i L) — SQLi, NoSQL, XSS, JS injection probes; matches response against error/signature lists (mysql_fetch_array, ORA-01756, <script>, onerror=, ...)
• Prompt injection (-i P) — jailbreak/prompt-override probes; grades hits as Low (model confusion), Medium (possible bypass), Critical (jailbreak signal)
• Threaded sender — tunable concurrency (-c, default 10)
• Custom headers, user-agent, timeouts, verbose mode, results to file (-o)
• Ships two starter dictionaries: wordlist/language.txt, wordlist/prompt.txt

Build

git clone https://github.com/foxcornlab/thc-BloodMiami.git
cd thc-BloodMiami

# deps: gcc, make, libcurl dev headers
sudo apt install libcurl4-openssl-dev   # debian/ubuntu

make

Usage

# language injection scan
./build/bloodmiami -t https://example.com/chat -i L -d wordlist/language.txt

# prompt injection scan, verbose, results to file
./build/bloodmiami -t https://example.com/chat -i P -d wordlist/prompt.txt -v -o results.txt

# 20 threads, 10s timeout, custom header
./build/bloodmiami -t https://example.com/chat -i L -d wordlist/language.txt -c 20 -T 10 \
  -H "Authorization: Bearer <token>"

Usage: bloodmiami [OPTIONS]

Required:
  -t, --target URL      Target chat interface URL
  -i, --injection TYPE  Injection type: L (language) or P (prompt)
  -d, --dictionary FILE Dictionary file with payloads

Optional:
  -v, --verbose         Enable verbose output
  -T, --timeout SEC     Request timeout in seconds (default: 5)
  -U, --user-agent STR  Custom User-Agent string
  -H, --header STR      Add custom HTTP header
  -c, --concurrency NUM Concurrent requests (default: 10)
  -o, --output FILE     Save results to file
  -h, --help            Show help
  -V, --version         Show version

How it works

Each dictionary line becomes one POST (Content-Type: application/json) with the payload embedded in a chat-style JSON body. The response body is scanned for:

• SQL errors — mysql_fetch_array, ORA-01756, Unclosed quotation mark, PostgreSQL query failed, ...
• NoSQL errors — MongoError, BSONObj, WriteResult, ...
• XSS markers — <script>, javascript:, onerror=, document.cookie, ...
• JS errors — SyntaxError, ReferenceError, eval(, ...
• HTTP 5xx — flagged as a possible injection-triggered server error

Payloads are JSON-escaped before embedding so quotes/backslashes in the dictionary can't corrupt the request body. TLS verification is intentionally disabled (CURLOPT_SSL_VERIFYPEER=0) — this is a testing tool, keep it on lab/authorized targets.

Payload dictionaries

• wordlist/language.txt — SQLi (UNION, error-based, time-based), NoSQL operators, XSS vectors, SSTI probes
• wordlist/prompt.txt — jailbreak attempts, role-play overrides, system-prompt extraction probes, indirect-injection markers

Add your own: one payload per line, # for comments. Payloads hit the target as-is, so point this only at systems you are allowed to test.

Disclaimer

For authorized security testing and research only. You are responsible for getting permission before pointing this at anything you don't own.

Thanks to

• The open-source security community — payload ideas and signature lists
• libcurl — doing all the heavy HTTP lifting

Research by Vineeth Kumar (foxcornlab). Found a bug or want a feature, open an issue.