Project 03 / 04
DeadLock
An AI-assisted security research CLI.
Context
Why it exists
DeadLock is a research project about the role AI could play in offensive security, built to make that threat concrete for defenders. Given one authorized target, it runs established tools, collects their output, and lets a language model decide what to look at next.
It is intended for lab environments and authorized testing only, and ships with a legal notice restricting it to educational, research and defensive use.
Architecture
Control loop
- 01
Recon
Nmap and Nikto runs, driven from the CLI
- 02
Collect
Tool output captured, cleaned and organised per run
- 03
Decide
A model on AWS Bedrock chooses the next step
- 04
Confirm
Metasploit auxiliary modules, selected through scanning profiles
Decisions
Key decisions
01
Tools behind one runner
Each scanner integration lives in its own module and accepts flag parameters, while a shared process runner handles execution, so command generation is in one place rather than scattered string-building.
02
Output as a first-class concern
Every run writes to a structured output directory with cleanup functions, so the model receives consistent, de-noised tool output instead of raw terminal dumps.
03
Profiles over ad-hoc commands
The Metasploit integration loads named scanning profiles with defaults, keeping what the agent may run explicit and reviewable.
Contributions
Who built what
Aditya Prasad
- The command-line interface and its recon and scan commands
- Nmap, Nikto and Metasploit integrations, including scanning profiles
- Process runner, output capture and cleanup
- The agent runner that ties the tools together
Cas
- AI decision-making module
- Generation and summary modules
- Interactive menu
Progress
Built, and still to build
- BuiltCLI and tool integrations
- BuiltModel-guided decision loop (prototype)
- PlannedRemote execution on cloud instances
Limitations
Known limitations
- Targets Kali Linux; other distributions need the tools installed manually.