Shape-Shifting Malicious Code in Software Backdoors via Language Models

Summary

Supply-chain attacks in open-source software are a notorious threat to security. Current defenses focus primarily on detecting backdoor functionality within the software. However, we show that seemingly benign documentation and configuration scripts can also serve as carriers of malicious code. To this end, we introduce an attack that uses large language models to encode a malicious payload into benign cover data. The resulting material appears natural and plausible to human reviewers, yet it can be easily reconstructed into its malicious form without access to a language model. Our evaluation demonstrates the efficacy of this approach in hiding code from audits. We argue that this form of shape-shifting code poses a notable risk and derive corresponding recommendations for software development.