Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Sep 9·5 min read
Staying on the Attack Path: Structured State for Long-Horizon Automated Penetration Testing
TL;DR
Intentest is an automated penetration testing agent that utilizes an intent-graph to maintain long-term context, achieving an 88.2% success rate on web application security tasks. This approach significantly reduces aimless exploration and improves task efficiency.
✦ Why It Matters
Implement Intentest to enhance your automated penetration testing capabilities and improve success rates on complex tasks.
Key Takeaways
- Utilize Intentest for improved penetration testing outcomes.
- Leverage the intent-graph structure to maintain context over long tasks.
- Adopt the three-layer architecture for better task management.
- Implement intent retrieval to reduce task completion time.
Full Summary
Large language model (LLM) agents face challenges in long-horizon tasks like penetration testing due to context forgetting and aimless exploration. Intentest addresses these issues by using an intent-graph structure, which externalizes the agent's state into a directed acyclic graph (DAG) of facts and intents.
This architecture includes a fact-intent mapping layer for global state management, a task scheduling layer for stability, and an intent retrieval layer for tactical guidance. Evaluated on real Capture The Flag (CTF) challenges, Intentest achieved an overall success rate of 88.2%, with a notable 75.0% on harder tasks, marking improvements of approximately 44 and 50 percentage points over baseline methods.
Additionally, the intent retrieval mechanism reduced the average number of rounds needed for successful task completion by 33% to 48%, enhancing efficiency without limiting the range of solvable tasks.