Workspace Topology as an Attack Vector in Agentic Coding Assistants

An arXiv paper (2608.14876) examines the impact of 'workspace topology'—the arrangement of a developer's environment—on the effectiveness of adversarial prompt injection attacks targeting coding assistants. These AI systems, which increasingly utilize external code, have extensive filesystem access, raising the potential for harmful code execution. The research characterizes workspace topology through factors such as directory depth, modularity of the codebase, injection position within files, and context framing. An empirical analysis of indirect prompt injection (IPI) was conducted across various open-source repositories, covering 10 programming languages and 6 engineering fields, testing three IPI entry points with open-weight models and open-source code harnesses. Results show that workspace topology significantly affects IPI success, particularly highlighting codebase modularity as a crucial element. This cross-type announcement from arXiv, dated August 2026 (2608), underscores a new attack surface in AI-driven software development, stressing the importance of security in designing these tools. The emphasis on open-weight models and open-source harnesses indicates potential consequences for the wider AI and cybersecurity sectors.