Skip to content
InfoResearchPreprintLLM-specific

Threat Landscape of Vibe Coding and the Road to Secure Autonomy

Published
Record updated
View JSON

Summary

This paper consolidates research on the security risks of vibe coding, where AI agents write, test and deploy code from natural language descriptions. It establishes a reference architecture, six levels of agent autonomy, and a threat model built on five trust boundaries, then proposes a taxonomy of twenty-one threats across nine classes. The authors find that risks grow with agents' operational authority, while most proposed defences rely on model judgement and address isolated threats.