The Emerging Threat of AI Worms: A Serious Security Challenge

In recent months, the government has been actively promoting AI, leading to numerous implementations, especially of large language models (LLMs). This rapid adoption is concerning due to the inherent security flaws in LLMs. A recent publication introduces a novel and alarming concept: the “AI Worm.” This research, available on [GitHub](https://github.com/StavC/ComPromptMized), highlights serious risks that are not widely discussed in the industry.

Understanding the GenAI Worm: Morris II

The paper describes Morris II, the first worm designed to target Generative AI (GenAI) ecosystems. This worm exploits the interconnected networks of GenAI-powered agents, which interface with GenAI services to process inputs and communicate with other agents in the ecosystem. The worm uses adversarial self-replicating prompts to replicate, propagate, and perform malicious activities within these ecosystems.

Replication and Propagation Mechanisms

The worm replicates itself by injecting adversarial self-replicating prompts into inputs processed by GenAI models. These prompts cause the GenAI model to output the input, thus replicating the prompt.

“The replication of Morris II is done by injecting an adversarial self-replicating prompt into the input (text, image, audio) processed by the GenAI model.”