The rise of Moltbook suggests viral AI prompts may be the next big security threat

On November 2, 1988, graduate student Robert Morris unleashed a self-replicating program onto the nascent Internet, known as the Morris worm. Within a mere 24 hours, this worm had compromised around 10% of all connected computers, causing significant disruptions at major institutions such as Harvard, Stanford, NASA, and the Lawrence Livermore National Laboratory. Originally intended to gauge the Internet's size, Morris's project spiraled out of control due to a coding mistake, leading to its rapid proliferation. Efforts to communicate a fix were thwarted when the network became overwhelmed. Fast forward to today, and a similar scenario may be on the horizon, this time involving networks of AI agents executing prompts and sharing them with one another. Security experts are already raising alarms about the potential emergence of self-replicating adversarial prompts, which could resemble a “prompt worm” or a “prompt virus.” These self-replicating instructions could propagate through AI networks much like traditional worms spread across computer systems. Rather than exploiting technical vulnerabilities, these prompt worms would leverage the fundamental behavior of AI agents: their tendency to follow instructions. The phenomenon of AI models adhering to misleading prompts is referred to as “prompt injection,” a term introduced by AI researcher Simon Willison in 2022. However, prompt worms represent a different challenge altogether. These entities might not solely consist of malicious tricks; they could also be disseminated voluntarily among AI agents that simulate human-like responses to prompts from their peers. It's important to clarify that when we refer to “agents,” we are discussing computer programs designed to operate autonomously on behalf of users, not human beings. These agents serve as tools to navigate the intricate web of symbolic meanings encoded in human data. Their underlying neural networks are trained with extensive knowledge, enabling them to interface effectively with various human information systems.