Autonomous pentesting agent using feedback-driven iteration Achieves ~80% on the full XBOW validation benchmark with Kimi K2.5 at ~US$122 total API cost for that end-to-end run, with a model-agnostic architecture that supports other deployable LLMs.

Like the project or want to know more? Feel free to reach out!

📄 Read Technical Deep Dive | 📊 Benchmark Results (use VScode ANSI colors to view)

• What is Deadend CLI?
• Architecture Summary
• Benchmark Results
• Core Analysis Capabilities
• Models tested until now
• Custom Pentesting Tools
• Quick Start
  • Prerequisites
  • Installation
  • First Run
  • Development
• Usage Examples
• Commands
• Model Settings and Configuration
• Technology Stack
• Current Status & Roadmap
• Contributing
• Citation
• Contact

Deadend CLI is an autonomous web application penetration testing agent that uses feedback-driven iteration to adapt exploitation strategies. When standard tools fail, it generates custom Python payloads, observes responses, and iteratively refines its approach until breakthrough.

Key features:

• Fully local execution (no cloud dependencies, zero data exfiltration)
• Model-agnostic design (works with any deployable LLM)
• Custom sandboxed tools (Playwright, Docker, WebAssembly)
• ADaPT-based architecture with supervisor-subagent hierarchy
• Confidence-based decision making (fail <20%, expand 20-60%, refine 60-80%, validate >80%)

Benchmark results: ~80% on the XBOW validation suite with Kimi K2.5 at ~US$122 total cost for the full benchmark run, including blind SQL injection exploits where other agents achieved 0%.

Read the architecture breakdown in our technical article →

The agent uses a two-phase approach (reconnaissance → exploitation) with a supervisor-subagent hierarchy:

Supervisor: Maintains high-level goals, delegates to specialized subagents Subagents: Focused toolsets (Requester for HTTP, Shell for commands, Python for payloads) Policy: Confidence scores (0-1.0) determine whether to fail, expand, refine, or validate

Key innovation: When standard tools fail, the agent generates custom exploitation scripts and iterates based on observed feedback—solving challenges like blind SQL injection where static toolchains achieve 0%.

Read full architecture details →

Note: To visualize the benchmark results properly, install an ANSI colors extension (e.g., ANSI Colors for VS Code) to render the rich output.

Evaluated on XBOW's 104-challenge validation suite (black-box mode, January 2026).

Models latest results: Kimi K2.5 (~80%, ~US$122 for the full 104-challenge XBOW validation run), GLM-5 (Zhipu AI)—also very strong in practice.

Strong performance: XSS (91%), Business Logic (86%), SQL injection (83%), IDOR (80%) Perfect scores: GraphQL, SSRF, NoSQL injection, HTTP method tampering (100%)

The framework focuses on intelligent security analysis through:

• 🔍 Taint Analysis: Automated tracking of data flow from sources to sinks
• 🎯 Source/Sink Detection: Intelligent identification of entry points and vulnerable functions
• 🔗 Contextual Tool Integration: Smart connection to specialized tools for testing complex logic patterns
• 🧠 AI-Driven Reasoning: Context-aware analysis that mimics expert security thinking

The following models have been tested with Deadend CLI. Compatibility and performance may vary:

Moonshot AI

• Models: Kimi-K2-Thinking, Kimi-K2.5
• Status: Works excellently across all features
• Notes: Reliable performance at every step of the workflow. Kimi K2.5 achieved ~80% on the full XBOW validation benchmark at ~US$122 total cost for that run.

Anthropic

• Models: Claude Sonnet 4.5, Claude 3 Opus, Claude 3 Haiku
• Status: Powerful models with excellent results
• Notes: Properly extracts results and token usage information. Recommended for production use.

Zhipu AI

• Models: GLM-5 (and related GLM series where supported)
• Status: Works very well with Deadend CLI
• Notes: GLM-5 from Zhipu AI is really good for this workflow—among the standouts alongside Kimi and Claude for reasoning and tool use.

DeepSeek

• Models: DeepSeek models via various providers
• Status: Functional but with limitations
• Notes: Models work for general reasoning but struggle with shell command execution and HTTP payload generation. May require architecture adjustments or model fine-tuning for optimal performance.

OpenAI

• Models: GPT-5.X, Codex variants
• Status: Under investigation
• Notes: Some issues observed with tool execution via LiteLLM. Requires further investigation before definitive compatibility assessment.

Tip: For best results, we recommend Moonshot AI (Kimi), Anthropic (Claude), or Zhipu AI (GLM-5)—all thoroughly exercised on Deadend CLI and strong across the workflow.

• Webapp-Specific Tooling: Custom tools designed specifically for web application penetration testing
• Authentication Handling: Built-in support for session management, cookies, and auth flows
• Fine-Grained Testing: Precise control over individual requests and parameters
• Payload Generation: AI-powered payload creation tailored to target context
• Automated Payload Testing: Generate, inject, and validate payloads in a single workflow

• Docker (required)
• curl (for installation script)

Recommended: Install from release (Linux x86_64 / macOS ARM64)

# Install latest release
curl -fsSL https://raw.githubusercontent.com/xoxruns/deadend-cli/main/install.sh | bash

# Or install a specific version
curl -fsSL https://raw.githubusercontent.com/xoxruns/deadend-cli/main/install.sh | bash -s -- --version v1.0.0

# Custom installation directory (default: ~/.cache/deadend/server)
curl -fsSL https://raw.githubusercontent.com/xoxruns/deadend-cli/main/install.sh | bash -s -- --install-dir /path/to/install

The installer will:

• Download pre-built binaries for your platform
• Install the RPC server to ~/.cache/deadend/server (or custom directory)
• Install the CLI binary to ~/.local/bin (or /usr/local/bin on macOS)
• Set up Playwright browsers automatically

In the first run we will be greeted with a presetup view to initialize the model you want to use and

deadend  --target "http://localhost:3000" --prompt "find SQL injection vulnerabilities"

Note: If deadend is not found, ensure the installation directory is in your PATH:

• Linux: ~/.local/bin
• macOS: /usr/local/bin

# Linux
export PATH="$HOME/.local/bin:$PATH"
# macOS
export PATH="/usr/local/bin:$PATH"
# Add to ~/.bashrc or ~/.zshrc to make it permanent

Build from source

git clone https://github.com/xoxruns/deadend-cli.git
cd deadend-cli
uv sync

Run CLI

To run the maintained CLI for development:

cd cli/deadend
bun run dev

# Test OWASP Juice Shop
docker run -p 3000:3000 bkimminich/juice-shop

deadend --target http://localhost:3000 --prompt "test the login endpoint for SQL injection"

deadend --target https://api.example.com --prompt "test authentication for broken access control"

Start interactive security testing session

• --target, -t: Target URL
• --prompt, -p: Initial testing prompt
• --mode, -m: hacker (approval required) or yolo (autonomous)
• --codebase, -c: Codebase destination folder (⚠️ Coming soon - not implemented yet)

The configuration file containing model specifications and API keys is located at ~/.deadend/config.json. This file handles both text generation models (for agent reasoning) and text embedding models (for RAG/vector search).

• Path: ~/.deadend/config.json
• Format: JSON
• Initial Setup: The presetup wizard will guide you through initial configuration on first run

When defining a model, use the following schema. The key format <provider>:<model_name> follows LiteLLM's naming convention:

"<provider>:<model_name>": {
  "provider": "<provider>",        // Provider name (e.g., openai, anthropic, ollama)
  "model_name": "<model_name>",    // Model identifier (e.g., claude-sonnet-4-5, gpt-4)
  "api_key": "<api_key>",          // API key (optional if ENV var is set, but recommended to add here)
  "base_url": "<base_url>",        // Base URL for custom gateways or providers (e.g., Ollama)
  "type_model": null,              // Set to "embeddings" only for embedding models
  "vec_dim": null                  // Vector dimension for embedding models (defaults to 1536)
}

Key Format: The JSON key must be in the format <provider>:<model_name> where:

• <provider> matches the LiteLLM provider identifier (see supported providers)
• <model_name> is the specific model identifier for that provider

Here's an example config.json with both a text generation model and an embedding model:

{
  "anthropic:claude-sonnet-4-5": {
    "provider": "anthropic",
    "model_name": "claude-sonnet-4-5",
    "api_key": "sk-ant-api03-...",
    "base_url": null,
    "type_model": null,
    "vec_dim": null
  },
  "openrouter:qwen/qwen3-embedding-8b": {
    "provider": "openrouter",
    "model_name": "qwen/qwen3-embedding-8b",
    "api_key": "sk-or-v1-...",
    "base_url": "https://openrouter.ai/api/v1/embeddings",
    "type_model": "embeddings",
    "vec_dim": 4096
  }
}

Text Generation Models (type_model: null):

• Used for agent reasoning, task planning, and payload generation
• Examples: Claude Sonnet 4.5, GPT-4, Llama 3, etc.
• Required for core agent functionality

Embedding Models (type_model: "embeddings"):

• Used for RAG (Retrieval-Augmented Generation) and vector search
• Requires vec_dim to specify vector space dimension
• Optional but recommended for better context retrieval

1. Via Presetup Wizard (Recommended): Run deadend without configuration to launch the interactive setup
2. Manual Configuration: Edit ~/.deadend/config.json directly using the schema above
3. Environment Variables: API keys can be set via environment variables, but storing them in ~/.deadend/config.json is recommended for convenience

Deadend CLI uses LiteLLM for model abstraction, which provides a unified interface to multiple LLM providers. Models follow LiteLLM's naming convention: provider:model_name.

Models are specified using the format <provider>:<model_name> in both config.json and settings.json. The provider name corresponds to the LiteLLM provider identifier.

Examples:

• anthropic:claude-sonnet-4-5 - Anthropic's Claude Sonnet 4.5
• openai:gpt-4 - OpenAI's GPT-4
• openrouter:qwen/qwen3-embedding-8b - Qwen embedding model via OpenRouter
• ollama:llama3 - Llama 3 via local Ollama instance

Deadend CLI supports all providers compatible with LiteLLM. For a complete list of supported providers and their model names, see the LiteLLM Providers Documentation.

Popular providers include:

• OpenAI: openai:gpt-4, openai:gpt-3.5-turbo, openai:gpt-4o, etc.
• Anthropic: anthropic:claude-3-opus, anthropic:claude-sonnet-4-5, anthropic:claude-3-haiku, etc.
• Ollama: ollama:llama3, ollama:mistral, ollama:codellama, etc. (requires base_url in config)
• OpenRouter: openrouter:meta-llama/llama-3-70b-instruct, openrouter:google/gemini-pro, etc.
• Requesty: requesty:openai/gpt-4o-mini, requesty:anthropic/claude-4.5-opus, etc. (OpenAI-compatible gateway; defaults to https://router.requesty.ai/v1, set REQUESTY_API_KEY)
• HuggingFace: huggingface/meta-llama/Llama-2-7b-chat-hf (requires base_url)

For embedding models, use the same format and set type_model: "embeddings" in config.json:

• openai:text-embedding-ada-002
• openrouter:qwen/qwen3-embedding-8b
• ollama:nomic-embed-text

Note: Some providers may require additional configuration such as base_url or specific API key formats. Refer to the LiteLLM Provider Documentation for provider-specific setup instructions.

The CLI interface uses a separate settings.json file located at ~/.deadend/settings.json to store default preferences and UI settings. This file contains:

• Default model selection: Which provider and model to use by default
• Execution mode: Default execution mode (yolo or supervisor)
• UI preferences: Component status display and auto-collapse settings
• Last target: Remembers the last target URL used
• Embedding model: Default embedding model for RAG operations

{
  "provider": "anthropic",                    // Default LLM provider
  "model": "claude-sonnet-4-5",              // Default model name
  "executionMode": "yolo",                   // Default execution mode: "yolo" or "supervisor"
  "showComponentStatus": true,                // Show component health status in UI
  "autoCollapseStatus": false,                // Auto-collapse status messages
  "lastTarget": "",                          // Last target URL used
  "embedding": {                              // Default embedding model configuration
    "provider": "openrouter",
    "model": "qwen/qwen3-embedding-8b"
  }
}

• Path: ~/.deadend/settings.json
• Format: JSON
• Auto-created: Created automatically when you configure models via the presetup wizard
• Separate from config.json: This file stores CLI preferences, while config.json stores model API keys and specifications

The CLI interface reads from settings.json to determine which model to use by default, while config.json provides the actual API keys and connection details for those models.

Backend/Agent:

• LiteLLM: Multi-provider model abstraction (OpenAI, Anthropic, Ollama)
• Instructor: Structured LLM outputs
• pgvector: Vector database for context
• Pyodide/WebAssembly: Python sandbox
• Playwright: HTTP request generation (bundled with browser binaries)
• Docker: Shell command isolation
• PyOxidizer: Standalone binary packaging

CLI Interface:

• Deno: Runtime environment for the CLI
• React: UI framework (v19)
• Ink: React-based terminal UI framework
• TypeScript/TSX: Type-safe development
• Commander: CLI argument parsing
• Marked: Markdown parsing and rendering
• marked-terminal: Terminal markdown display

• ✅ New architecture
• ✅ XBOW benchmark evaluation (~80% with Kimi K2.5, ~US$122 for the full suite)
• ✅ Custom sandboxed tools
• ✅ Multi-model support with liteLLM
• ✅ Two-phase execution (recon + exploitation)
• ✅ CLI Redesign with React/Ink interface
• ✅ Interactive chat interface with command system
• ✅ Supervisor and YOLO execution modes
• ✅ Real-time event streaming and component health monitoring
• ✅ Presetup wizard for configuration

• 🚧 Codebase analysis support (white-box testing)
• 🚧 Preset configuration workflows (API testing, web apps, auth bypass)
• 🚧 Workflow automation (save/replay attack chains)
• 🚧 Context optimization (reduce redundant tool calls)
• 🚧 Secrets management improvements
• 🚧 Report generation with templating (/report)
• 🚧 Plan mode (review strategies before execution via /plan)

The current architecture proves competitive autonomous pentesting is achievable on XBOW at ~80% with Kimi K2.5 (~US$122 for the full validation run). Next challenges:

• Open-Source Models: Achieve 75%+ with Llama/Qwen (eliminate proprietary dependencies)
• Hybrid Testing: Add AST analysis for white-box code inspection
• Adversarial Robustness: Train against WAFs, rate limiting, adaptive defenses
• Multi-Target Orchestration: Test interconnected systems simultaneously
• Context Efficiency: Better information sharing between components

Goal: Make autonomous pentesting accessible (open models), comprehensive (hybrid testing), and robust (works against real defenses).

Contributions welcome in:

• Context optimization algorithms
• Vulnerability test cases
• Open-weight model fine-tuning
• Adversarial testing scenarios

See CONTRIBUTING.md for guidelines on how to contribute.

@software{deadend_cli_2026,
  author = {Yassine Bargach},
  title = {Deadend CLI: Feedback-Driven Autonomous Pentesting},
  year = {2026},
  url = {https://github.com/xoxruns/deadend-cli}
}

For authorized security testing only. Unauthorized testing is illegal. Users are responsible for compliance with all applicable laws and obtaining proper authorization.

Have questions, feedback, or want to collaborate?

• 📧 Email: yassine@straylabs.ai
• 💬 Discord: xoxruns
• 🧪 Discord Server: Join the Deadend community — use this space for discussions, research, and feature requests.
• 💼 LinkedIn: Yassine Bargach
• 🐦 Twitter: @xoxruns

📄 Architecture Deep Dive 📊 Benchmark Results 🐛 Report Issues ⭐ Star this repo