The authors collected 83 smart contract audit skills from public sources and tested them across seven agent–model configurations on the EVMBench benchmark. They found that skills are typically lightweight and diverse in design, covering many vulnerability types but with an imbalanced distribution. Skill effectiveness depended more on the backend model than the ; Codex/GPT-5.5 achieved the largest gains, improving detection score by 22.8% and captured award by 43.2%. However, consistent activation of skills was a major bottleneck, with only one configuration triggering them reliably. The study provides the first empirical characterization of audit skills and releases the skill corpus for future research.
Reusable skills improve smart-contract audits, but model choice dominates
Across 83 audit skills and seven agent–model configurations, performance gains depended chiefly on the underlying model and whether the agent activated the skill.
Top University
Cuifeng Gao · Juantao Zhong · Jiachi Chen · Shuai Wang · Daoyuan Wu
Lingnan University · The Hong Kong Polytechnic University · State Key Laboratory of Blockchain and Data Security · Zhejiang University · Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security
Research Digest··2 min read
Thread:Skill Selection for Agents
In one line
Model capability, not agent harness, determines skill effectiveness in smart contract auditing; triggering is a bottleneck.
Before you read this
Three ideas this paper builds onagent skillsLearned capabilities that enable AI agents to perform specialized tasks with improved effectiveness.118 papers →audit workflowA systematic process for verifying data integrity, model behavior, and evaluation validity through checks for errors, confounds, and hidden artifacts.108 papers →agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.176 papers →
Read it as
The authors collected smart-contract auditing skills used in the wild and evaluated them on EVMBench while analyzing how they changed agent execution.
The authors systematically collected 83 reusable smart-contract audit skills, examining their structure, encoded security knowledge, workflows, and tool dependencies.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.