Pwnagent: a knowledge-guided multi-agent system for automatic exploit generation

An LLM-driven multi-agent framework named PwnAgent was developed for end-to-end exploit generation by combining offensive domain knowledge with active runtime introspection, aiming to bridge the gap between static vulnerability analysis and dynamic memory behavior. The system employs a hierarchical knowledge base structured around four layers—Foundational Principles ($L1$), Reasoning Rule Chains ($L2$), Reference Trajectories ($L3$), and Empirical Pitfalls ($L4$)—to guide multi-stage exploit reasoning. PwnAgent’s architecture features a Reasoning Layer centered on a MainOrchestrator and an Execution Layer comprising specialized agents (ProgramAnalyzer, MeasurementExpert, ExploitCrafter) interacting via a unified Environment Interaction Bus. The system's core innovation is the structured knowledge approach, utilizing cognitive sharding to assign role-specific segments of this knowledge to specialized agents to manage context effectively. Evaluation involved a 66-task pwn benchmark of Linux ELF binaries and stack-oriented tasks. PwnAgent achieved a 62.12% end-to-end success rate on this benchmark, demonstrating a 30.30 percentage-point gain over the PwnGPT baseline (31.82%), showing that execution-grounded measurement and structured knowledge guidance improve LLM-based exploit generation.

The work addresses the fundamental Logical-Physical Gap in LLM-based exploitation by architecting a closed-loop system that forces reasoning to be grounded in execution. The insight that static semantic reasoning fails to account for dynamic physical invariants—like compiler alignment padding or runtime memory states—is critical. The framework's success stems from decomposing this gap into three challenges (C1: introspection deficit, C2: knowledge representation gap, C3: feedback ambiguity) and designing specialized agents to address them sequentially. The Cognitive Sharding mechanism is a sophisticated structural choice; it sacrifices the potential for maximal context sharing across all stages in favor of deterministic, low-latency reasoning by strictly bounding cognitive scope via static role assignments. This suggests that successful agentic systems in complex domains are less about maximal information flow and more about enforced, expert-aligned procedural decomposition. The result, a performance gain, confirms that incorporating structured, execution-grounded constraints fundamentally changes the quality of autonomous output, shifting the focus from pattern matching to verifiable operational adherence. The implication for future work is whether this structure can generalize beyond pwn challenges to other domains where dynamic ground truth is equally elusive.