BFSI insights

AI Agentic Vulnerability Injection And Transformation with Optimized Reasoning

Published 6 Nov 2025 · arXiv · Amine Lbath
arXiv preview

Overview

The paper presents AVIATOR, an AI-agentic workflow designed to inject realistic software vulnerabilities, addressing limitations in existing datasets. It achieves high success rates in generating diverse and accurate vulnerability datasets.

Key Insights

  • Insight: AVIATOR achieves 91%-95% injection success rates.
    • Evidence: Evaluations across three benchmarks.
    • Verifiable: Yes
  • Insight: AVIATOR uses a modular approach with specialized AI agents and traditional tools.
    • Evidence: Combines semantic analysis, LoRA-based fine-tuning, and static analysis.
    • Verifiable: Yes

BFSI Relevance

  • Why Relevant: Enhances cybersecurity measures by improving vulnerability detection systems.
  • Primary Sector: Financial Services
  • Subsectors: Cybersecurity
  • Actionable Implications:
    • Implement AVIATOR to enhance vulnerability detection.
    • Use generated datasets for training AI models in cybersecurity.
researcher peer-reviewed-paper global
Open on arXiv