Oubliette Dungeon
Automated adversarial testing for LLM applications. 72 attack scenarios (57-scenario core library + 15 Crescendo multi-turn), 10 categories. Know your blind spots before attackers do.
pip install oubliette-dungeon 72 Attack Scenarios, 10 Categories
57-scenario core library + 15 Crescendo multi-turn scenarios. Together they cover all 10 OWASP LLM Top 10 categories (9 in the core 57). Most scenarios also carry a MITRE ATLAS mapping.
Features
Everything you need to red team your LLM applications.
72 Attack Scenarios
72 YAML-defined attack scenarios in two sets: a 57-scenario core library and 15 Crescendo multi-turn scenarios. Mapped to OWASP LLM Top 10 and MITRE ATLAS (v2026.06). Covers prompt injection, jailbreak, information extraction, context manipulation, and more.
Multi-Provider Comparison
Run the same attack suite against OpenAI, Anthropic, Ollama, Azure, and more. Compare detection rates side by side.
React Dashboard
Interactive dashboard with real-time results, trend analysis, provider comparison charts, and scenario drill-down.
CLI + API
Click-based CLI for scripting and CI/CD. REST API for programmatic access. JSON and HTML report generation.
Tool Integrations
Optional adapters for PyRIT, DeepTeam, Garak, and AIX. Use industry-standard tools alongside Dungeon scenarios.
Scheduled Campaigns
Schedule recurring red team campaigns. Track detection rate trends over time. Get alerts when regressions are detected.
Quick Start
Three commands to your first red team campaign.
pip install oubliette-dungeon oubliette-dungeon run --target http://localhost:5000/api/chat oubliette-dungeon compare --providers openai,anthropic,ollama MCP Server Integration
Run red team scenarios directly from Claude Desktop or any MCP-compatible client. No CLI needed.
list_scenarios
Browse and filter the 57-scenario core library by category, difficulty, or keyword.
run_scenario
Execute a single scenario against any target URL with full result analysis.
run_category
Run all scenarios in a category and get aggregated detection rate metrics.
get_results
Query historical test results from the built-in results database.
get_metrics
Compute detection rate, pass@k, risk density, and other security metrics.
export_report
Generate JSON or HTML reports from test campaign results.
MCP servers available on request: contact info@oubliettesecurity.com
Pairs with Oubliette Shield
Dungeon attacks. Shield defends. Run Dungeon against a Shield-protected endpoint to measure your detection rate and find gaps.
Learn About ShieldStart Testing Your AI Today
Free and open source. Enterprise support available.