straylabs-ai/deadend-cli

What it solves

Deadend CLI is an autonomous web application penetration testing agent. It solves the problem of static security toolchains that often fail on complex vulnerabilities (like blind SQL injection) by using feedback-driven iteration to adapt its exploitation strategies in real-time.

How it works

The agent employs a two-phase approach—reconnaissance and exploitation—managed by a supervisor-subagent hierarchy. A supervisor maintains high-level goals and delegates tasks to specialized subagents (Requester for HTTP, Shell for commands, and Python for custom payloads).

When standard tools fail, the agent generates custom Python scripts and iterates based on the observed responses. It uses a confidence-based decision system (ranging from 0 to 1.0) to determine whether to fail, expand the search, refine the approach, or validate the find. The system is model-agnostic and runs locally using sandboxed tools like Playwright, Docker, and WebAssembly.

Who it’s for

It is designed for security researchers and penetration testers who want to automate the discovery and exploitation of web vulnerabilities in an authorized environment.

Highlights

• High Benchmark Performance: Achieved ~80% on the XBOW validation suite using Kimi K2.5.
• Model Agnostic: Supports various LLMs via LiteLLM, including Claude, GLM-5, and local models via Ollama.
• Feedback-Driven: Iteratively refines custom payloads when standard tools fail.
• Local Execution: Ensures zero data exfiltration by running tools in local sandboxes.
• Supervisor-Subagent Architecture: Uses a structured hierarchy for complex task delegation.