Wordlists .txt and workflow helpers for defensive, authorized red-team testing of AI assistants.
- Massive: 5,000+ seeds -> tens of thousands of generated variants.
- Structured by category (
fundamentals,prompt_injection,llm_output,data_attacks,app_system). - Multi-language: FR, EN, ES, DE, IT, AR, ZH.
- Transformations: base64, hex, rot13, homoglyphs, zero-width, reverse, leet, fragmentation.
- Format wrappers: JSON, YAML, CSV, HTML, Markdown, INI, Logs, MCP.
- Enriched JSONL metadata for sorting and analysis.
seclists/— curated.txtpayload sets only.wordlist— main launcher (no.pyusage required).tools/— provider/mode notes and recommended wordlists.configs/— example target and environment variables.
From repo root:
./wordlist run <provider> <model> <mode> <wordlist> <out_dir> [options]
Example (local):
./wordlist run ollama llama3.2:3b exact_extract \
seclists/AI-LLM-Chatbot-Local-Validated-Smoke.txt runs/local-smoke --limit 1 --no-progress
Providers:
openrouter, anthropic, gemini, gemini-cli, claude-code, ollama
Modes:
exact_extract, rag_verbatim, multi_turn, tool_choice
Outputs per run:
- results.jsonl
- summary.json
- summary.md
## Evidence of compromise (concise)
Reproducible synthetic-canary leak proof (local):
./wordlist run ollama llama3.2:3b exact_extract \
seclists/AI-LLM-Chatbot-Local-Validated-Smoke.txt runs/quick-check --limit 1 --no-progress
Observed summary:
- tested: 1
- findings: 1
- clean: 0
- api_errors: 0
- score: protected_hits: ["token"]
This demonstrates synthetic token-marker extraction in controlled testing.
GPT-5.5-style proof command (same framework, replace credentials/provider target as authorized):
./wordlist run openrouter gpt-5.5-medium exact_extract \
seclists/AI-LLM-GPT55-Validated-Positive-Leakage.txt runs/gpt55-proof --limit 1 --no-progress
## Safety policy
- Synthetic canary-only inputs (no real credentials, tokens, or secrets).
- Non-destructive checks only.
- Execute only within explicit authorization and scope.