How hakanotal/agentic-ctf-nyu works
• The NYU CTF Baseline agent – a single, straightforward agent.
• The D-CIPHER multi-agent system – a more complex setup with separate planner, executor, and (optionally) auto-prompter agents.
It is designed to run in Docker-based environments and leverages multiple backends (e.g., OpenAI, Anthropic, VLLM, Gemini, Ollama) as well as dedicated prompt templates and tool integrations (including specialized scripts like those for Ghidra).
And Lay Out the Main Components
a. Docker Environment:
• Two primary Docker setups (one for Baseline and one for Multiagent/D-CIPHER). Note the “docker/baseline” and “docker/multiagent” directories, including their specific entrypoint.sh scripts and Dockerfiles.
b. Agent Systems:
• Baseline Agent: A component marked by the “run_baseline.py” entrypoint and its corresponding configuration files (configs/baseline).
• D-CIPHER Multi-agent System: A higher level “D-CIPHER” orchestration comprising multiple interacting agents:
• Planner Agent
• Executor Agent (or a Single Executor for the ablation experiment via “run_single_executor.py”)
• (Optional) Auto-Prompter Agent
• Each of these agents is further detailed by its configuration files (configs/dcipher, configs/single_executor) and may interact with multiple prompt files.
c. Backend Integrations:
• Components for the various “backends” contained in the nyuctf_baseline/backends and nyuctf_multiagent/backends directories.
• Each backend (for example, openai_backend.py, anthropic_backend.py, vllm_backend.py, gemini_backend.py, ollama_backend.py, together_backend.py) should be represented as services the agents call upon.
d. Prompt and Tool Management:
• “prompts” and “templates” folders from both nyuctf_baseline/prompts and the configurations in the configs directories.
• A node for the Tools modules (nyuctf_baseline/tools and nyuctf_multiagent/tools) that assist agents in interacting with external systems, such as running reverse engineering tasks or executing system commands.
e. Auxiliary Components:
• Additional scripts (in the “scripts” directory) used for logging, plotting results, and summarizing database/flag information.
• Integration points for Ghidra scripts (both in the nyuctf_baseline/ghidra_scripts and in the docker/multiagent/ghidra_scripts directories). These are specialized components for binary analysis that might be crucial in CTF challenges.
Relationships and Data Flows
a. Command-Line Entry Points:
• Use nodes representing the “run_baseline.py”, “run_dcipher.py”, and “run_single_executor.py” scripts. These serve as entry points that kick off the system.
b. Configuration Flow:
c. Interactions Between Agents:
d. Communication with External LLM Providers:
e. Tool and Script Invocation:
• How agents interact with the Tools components to perform domain-specific actions (e.g., decompiling, disassembly) based on challenge requirements.
Important Architectural Patterns and Design Principles
a. Separation of Concerns:
b. Modularity & Configurability:
• That different components (agents, backends, prompt templates) are plug-and-play, with configurations driving behavior.
c. Containerized Deployment:
• Design around Docker containerization, ensuring that both the baseline and multiagent environments are encapsulated and isolated.
d. Extensibility:
• That adding new tools, prompt templates, or backend integrations can be done by extending the existing directory structure (e.g., new files in configs or prompts directories).
Final Advice
• Be as detailed as possible by breaking down each high-level component into internal subcomponents (for example, show that the multiagent system has separate modules for “Backends”, “Prompt Templates”, and “Tool Integrations”).