AgenticRepair: Multi-Faceted Context Engineering for Automated Vulnerability Repair
A novel framework named AgenticRepair seeks to enhance the automated repair of vulnerabilities by creating a more comprehensive program context. This methodology, outlined in a paper available on arXiv (ID 2607.29422), tackles three significant deficiencies in current agentic AI techniques: code-structure context (including cross-file data flows and memory operation patterns), runtime-execution context (focusing on crash semantics and memory origins), and commit-history context (examining the introduction of fragile code patterns). AgenticRepair employs three specialized LLM subagents to develop these contexts for integration into the repair process. The paper emphasizes that vulnerability repair necessitates more context than typical bug fixing, as security engineers typically gather such information. The framework aims to streamline the time and effort involved in addressing security flaws identified in vulnerability triage reports. This paper was presented as a cross-type submission and can be accessed via the provided URL.
Key facts
- AgenticRepair is a framework for automated vulnerability repair.
- It addresses three gaps: code-structure, runtime-execution, and commit-history context.
- The framework uses three specialized LLM subagents to engineer contexts.
- The paper is on arXiv with ID 2607.29422.
- The announcement type is 'cross'.
- The goal is to reduce time and effort for patching security flaws.
- Existing agentic approaches do not engineer the required context.
- The paper was published on arXiv.
Entities
Institutions
- arXiv