How to Use PentAGI for Autonomous AI Red Teaming

PentAGI is a fully autonomous AI agents system by vxcontrol that performs complex penetration testing tasks using multiple coordinated AI agents that talk to each other to hack a target. No human input required after initialization. While Eigent coordinates multi-agent teams for general automation, PentAGI specializes in security testing with dedicated orchestrator, researcher, developer, and executor agents.

• Project link: github.com/vxcontrol/pentagi

How PentAGI Works

PentAGI runs a team of specialized AI agents inside an isolated Docker environment. Three agent types collaborate to execute a penetration test from reconnaissance to exploitation.

Agent	Role	Responsibility
Orchestrator	Commander	Assigns tasks, coordinates workflow, manages state
Researcher	Intelligence	Scans targets, gathers OSINT, identifies vulnerabilities
Developer	Planner	Designs attack chains, selects exploits, plans execution
Executor	Operator	Runs tools, executes payloads, documents results

The flow follows a classic penetration testing lifecycle:

Orchestrator assigns target
     ↓
Researcher performs reconnaissance (nmap, OSINT, web scraping)
     ↓
Developer analyzes findings and plans attack vectors
     ↓
Executor runs exploits (Metasploit, sqlmap, custom payloads)
     ↓
Results stored in vector database for future reference
     ↓
Chain summarization prevents token overflow across long sessions

Architecture

PentAGI is built as a microservices architecture running on Docker Compose. Each component is horizontally scalable. For production deployments, NVIDIA NemoClaw hardens agent runtimes against escape attempts that could leak sensitive data.

Core Components

Component	Technology	Purpose
Frontend UI	React + TypeScript	Web dashboard for system management
Backend API	Go + GraphQL	Core logic, agent orchestration, data layer
Vector Store	PostgreSQL + pgvector	Persistent storage of commands, outputs, embeddings
AI Agents	Multi-agent Go system	Specialized agent orchestration
Knowledge Graph	Graphiti + Neo4j	Semantic relationship tracking
Security Tools	20+ tools sandboxed	nmap, Metasploit, sqlmap, etc.
Web Scraper	Isolated browser	OSINT gathering from web sources

Memory System

PentAGI implements a three-tier memory architecture that persists knowledge across sessions:

• Long-term Memory: Vector embeddings of past findings, domain expertise, and tool usage patterns stored in pgvector.
• Working Memory: Current task state, active goals, and system resources tracked in real time.
• Episodic Memory: Historical actions, outcomes, and success patterns that inform future decisions.

The chain summarization system prevents token limits from being exceeded during long penetration tests. It selectively summarizes older messages while maintaining conversation coherence, using configurable parameters like SUMMARIZER_PRESERVE_LAST and SUMMARIZER_MAX_QA_SECTIONS.

Quick Start

Prerequisites

• Docker and Docker Compose installed
• An LLM API key (OpenAI, Anthropic, Ollama, or any supported provider)
• A target system you have permission to test

Deployment

git clone https://github.com/vxcontrol/pentagi.git
cd pentagi
cp .env.example .env
# Edit .env with your LLM API keys and target configuration
docker compose up -d

The web UI will be available at https://localhost:8443 after startup.

API Access

PentAGI exposes both REST and GraphQL APIs with Bearer token authentication for automation and integration:

curl -X POST https://localhost:8443/api/v1/flow \
  -H "Authorization: Bearer YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"target": "https://target-system.com", "flow_type": "full_audit"}'

LLM Provider Support

PentAGI supports 10+ LLM providers and aggregators, making it flexible for any deployment:

Provider	Type	Configuration
OpenAI	Cloud	API key via environment variable
Anthropic	Cloud	API key via environment variable
Google Gemini	Cloud	API key via environment variable
AWS Bedrock	Cloud	AWS credentials
Ollama	Local	Endpoint URL
DeepSeek	Cloud	API key
GLM (Z.AI)	Cloud	API key
Kimi (Moonshot)	Cloud	API key
Qwen (Alibaba)	Cloud	API key
OpenRouter	Aggregator	API key
DeepInfra	Aggregator	API key

For production local deployments, the project recommends vLLM with Qwen3.5-27B-FP8 for self-hosted inference.

A quick detour we think you'll thank us for: How to Patch GIMP to Look Like Photoshop with PhotoGIMP and How to Use Dexter for Autonomous Financial Research.

Using Ollama (Local)

# In your .env file
LLM_TYPE=ollama
OLLAMA_ENDPOINT=http://host.docker.internal:11434
LLM_MODEL=llama3.3:70b

[!TIP]

For best results with local models, PentAGI recommends at least a 70B parameter model. The chain summarization system helps smaller models by keeping context within their effective window.

Key Features

• Secure and Isolated: All operations run in sandboxed Docker environments with complete isolation from the host.
• Fully Autonomous: AI-powered agents automatically determine and execute penetration testing steps with optional execution monitoring.
• Professional Tool Suite: Built-in access to 20+ professional security tools including nmap, Metasploit, sqlmap, and more.
• Smart Memory: Long-term storage of research results and successful approaches for future tests.
• Knowledge Graph: Graphiti-powered Neo4j integration for semantic relationship tracking.
• Web Intelligence: Built-in browser scraper for gathering latest information from web sources.
• External Search: Integration with Tavily, Perplexity, DuckDuckGo, Google Custom Search, Sploitus, and Searxng.
• Comprehensive Monitoring: Grafana dashboards, Prometheus metrics, Jaeger distributed tracing, and Loki log aggregation.
• Detailed Reporting: Generates thorough vulnerability reports with exploitation guides.
• Langfuse Integration: LLM observability with ClickHouse analytics and Redis caching.

Langfuse Integration Details

PentAGI integrates Langfuse for LLM observability. This provides:

• Trace-level monitoring of every agent decision
• Cost tracking per task and per model
• Performance metrics for response times and token usage
• ClickHouse-backed analytics for historical queries
• Redis caching for rate limiting and prompt caching
• MinIO S3 storage for trace artifacts

Enable it by setting LANGFUSE_ENABLED=true in your .env file.

What the Community Is Saying

The Threads post ignited a heated discussion about the future of cybersecurity:

“It’s already like this. I get all excited when I catch on to something only to get past the AI and it’s like I burst a bubble. Expert technique and then suddenly I’m dealing with a script kiddie. I had one fall for a canary token.” — @cheezecake2459

“Yeah there are a few programs like this in the works. Honestly I do think it’s going to be the future. Humans are not going to be able to keep up with this.” — @0xtoxsec

“Seeing as we don’t have AGI yet, the name is misleading.” — @c_h_a_r_t_r_e_u_s_e

The sentiment captures the divide: PentAGI is not true AGI, but its autonomous multi-agent approach already outperforms manual testing in speed and coverage. The tool is designed for cybersecurity professionals and researchers who need a powerful, flexible solution for conducting penetration tests at scale. Whether you see it as a force multiplier or a disruptor, the age of AI-driven red teaming has arrived.

If you enjoy articles about top GitHub repositories like this, don’t forget to subscribe to Technolati.com.