Protocol-Level Attacks on Agentic Commerce Platforms: Taxonomy and Defense (arxiv.org)

A recent study has unveiled critical security vulnerabilities in agentic commerce platforms, which allow AI agents to autonomously perform tasks like discovering services and processing payments. Researchers have identified 33 structural vulnerabilities in the protocols between these agents and commerce services, demonstrating that these flaws lead to a 100% attack-success rate, irrespective of the AI model's sophistication. This research shifts the focus from model-dependent security issues, such as prompt injection, to protocol-level weaknesses that are systemic across various platforms. The significance of this work lies in its potential to reshape how security is approached in AI commerce. By introducing a taxonomy that categorizes these protocol-level attacks, the study emphasizes the need for robust defenses that operate at the protocol layer, leading to the development of AIP-Bench, a benchmark for agentic commerce security. Additionally, the proposed defense mechanism, PCAT, effectively reduces the attack-success rate for most structural classes to zero, suggesting that securing agentic commerce platforms requires a comprehensive understanding of both the model and protocol layers, opening avenues for more resilient systems in this emerging domain.