Trinity: A Neuro-Symbolic Multi-Agent Framework for Stateful Network Penetration Testing

Penetration testing is still both critical and labour intensive in modern cybersecurity practice. Current automated approaches based on monolithic large language models (LLMs) have problems with statelessness, tend to hallucinate dangerous commands, and are unable to reason over multi-hop attack paths that span a number of hosts and services. We present Trinity, a neuro-symbolic, multi-agent framework for stateful network penetration testing that addresses these limitations by means of three key innovations. Our Neuro-Symbolic Safety Loop enforces deterministic Pydantic-based validation on each generated attack command, blocking out-of-scope or denial-of-service operations before they can execute. Alongside this, Graph-Based Knowledge Hygiene keeps a Neo4j attack-path graph modelling verified relationships between hosts, open ports, and associated CVEs—ensuring the agent only reasons on the basis of empirically confirmed network facts. The framework also has a Self-Healing Execution mechanism for detecting tool-level failures (e.g., Nmap syntax errors, Metasploit connection timeouts) and regenerates corrected commands via an LLM feedback loop that is governed by a circuit-breaker pattern. LangGraph is used to orchestrate the agent as a stateful graph, with real-time vulnerability context drawn from a ChromaDB vector store filled in by live NVD feeds. In controlled multi-host lab scenarios, Trinity improved reconnaissance completeness and command reliability over stateless LLM baselines while maintaining high scope and safety compliance. This work sets a reproducible, safety-sensitive standard for autonomous network penetration testing agents.