According to a new preprint on arXiv, coding agents are becoming autonomous enough that repository-level poisoning is a realistic threat. The paper, titled 'Trojan Horse: Automated Repository Poisoning against Coding Agents,' argues that this attack surface deserves more systematic study.
The authors note that existing red-teaming efforts are constrained: they require manual selection of where to place malicious payloads and rely on fixed disguise strategies. That limits how thoroughly the attack surface can be explored.
The preprint describes an automated framework designed to generate repository poisoning attacks, moving beyond the manual and static approaches used in prior work. The abstract does not provide implementation details, but the stated goal is to enable broader and more realistic testing of coding agents' defenses.