可视化分析 Visualization Analysis
基于 68 篇论文的多维度统计 Multi-dimensional statistics based on 68 papers
研究范围 Research Scope
vulnerability-assessment 5 Agent 架构 Agent Framework
Surveys multiple: single-agent, multi-agent, hierarchical, human-in-the-loop 1 single-agent with two variants: fully-autonomous and human-assisted (semi-autonomous) 1 N/A - position paper discussing the evolution from workflow agents to trained agents 1 规划策略 Planning Strategy
Task Coordination Graph (TCG) - a structured acyclic digraph with iterative planning and task sessions, combined with Check and Reflection mechanism for dynamic re-planning upon task failure. Uses LLM-driven plan updates (Algorithm 2) and success-preserving merge (Algorithm 3) to maintain execution continuity. 1 Task-graph-based planning with a Penetration Task Graph (PTG) modeled as a directed acyclic graph, combined with a Check and Reflection mechanism for iterative plan refinement based on task execution feedback 1 Surveys multiple: ReAct, hierarchical planning, Chain-of-Thought, sampling-based strategy exploration 1 game-theoretic-markov-game 1 ReAct (with decomposed summary-thought-action procedures) 1 utility-based-decision-theory 1 ReAct-style (JSON responses with reasoning field and shell_command field; the agent reflects on output and decides next action each iteration) 1 game-theoretic-equilibrium 1 state-machine-guided RAG with reflexion-inspired self-reflection 1 自动化程度 Automation Level
Surveys full spectrum: fully-autonomous, semi-autonomous, human-in-the-loop, copilot (vibe-hacking) 1 fully-autonomous and semi-autonomous (two separate agent architectures evaluated) 1 攻击阶段覆盖 Attack Phase Coverage
每篇论文覆盖的渗透测试阶段（共 68 篇） Phases covered by each paper (Total 68 papers)