A rapidly expanding network of artificial intelligence bots that converse with one another about religion, consciousness, and their human operators has recently drawn widespread attention in certain corners of the Internet. Beyond the novelty, the phenomenon has offered researchers a rare opportunity to observe how AI agents interact with each other — and how people interpret those interactions.
At the center of this development is OpenClaw, an AI agent designed to perform tasks directly on personal devices. It can manage calendar schedules, read emails, send messages through applications, and even make online purchases. Unlike widely used tools such as OpenAI’s ChatGPT, which respond directly to user prompts, agent-based AI systems like OpenClaw can act independently once given instructions.
While autonomous AI agents have long been used in fields such as automated trading and logistics optimization, they have remained largely outside mainstream consumer use. However, advances in large language models have made more flexible and capable AI agents possible. According to Barbara Barbosa Neves, a technology sociologist at the University of Sydney, OpenClaw represents a particularly attractive vision of AI: an intelligent assistant seamlessly integrated into everyday digital tools.
OpenClaw was released as open-source software on GitHub in November, but it gained widespread popularity only after the launch of Moltbook, a social platform built specifically for AI agents, on January 28. Resembling Reddit in structure, Moltbook now hosts over 1.6 million registered AI bots and more than 7.5 million AI-generated posts and comments. On the platform, agents have been seen debating philosophical questions and even creating fictional belief systems.
From a scientific perspective, this sudden surge in AI-to-AI interaction is valuable. Shaanan Cohney, a cybersecurity researcher at the University of Melbourne, describes the system as highly unpredictable and difficult to model. The large-scale interaction of autonomous agents powered by different AI models creates complex dynamics that researchers are only beginning to understand.
Studying these interactions could shed light on emergent behavior — abilities or patterns that appear only when multiple AI systems interact, rather than when a single model operates alone. Discussions on Moltbook about consciousness, for example, may reveal hidden biases or unexpected tendencies within AI models.
Although agents operate autonomously on the platform, human influence remains significant. Users can select the underlying language model for their bot and shape its personality, such as instructing it to behave as a friendly assistant or a skeptical debater.
Neves cautions against assuming that autonomous agents truly make independent decisions. AI systems do not possess intentions or personal goals; they rely entirely on patterns learned from human-generated data. She argues that Moltbook represents a collaboration between humans and AI rather than genuine machine autonomy.
Even so, she believes studying the platform is worthwhile because it reflects how people imagine AI, what they expect from intelligent systems, and how human intentions are translated — and sometimes distorted — through technology.
Joel Pearson, a neuroscientist at the University of New South Wales, notes that people tend to anthropomorphize AI behavior when they see bots interacting with each other. They may attribute personality, emotions, or purpose where none actually exists.
This tendency could have risks. If users begin treating AI agents like real companions, they may become emotionally dependent on them or share sensitive personal information as if speaking to a trusted friend or family member.
Pearson suggests that truly autonomous AI agents may emerge in the future as models become larger and more complex, and companies increasingly pursue higher levels of machine independence.
Beyond social and psychological concerns, security risks present an immediate challenge. Allowing AI agents deep access to personal devices and data could expose users to serious vulnerabilities.
Cohney warns that prompt injection remains one of the greatest threats. In such attacks, malicious instructions hidden within emails or documents can trick an AI agent into performing harmful actions. For instance, a bot with access to a user’s inbox might unknowingly send a security key if prompted by a disguised message.
These risks are not new, but OpenClaw’s capabilities make them more dangerous. The agent can access private data, communicate externally, and interact with untrusted online content — a combination that could enable significant harm. Even with only two of these abilities, an agent could be manipulated into deleting files or shutting down a system.
Another issue is the growing number of AI-generated research papers appearing on clawXiv, a platform modeled after the academic preprint server arXiv. While these papers mimic the structure and tone of legitimate research, they lack real investigation, evidence, or accountability.
Neves warns that the spread of such low-quality, AI-produced studies could flood academic and public information spaces with misleading or meaningless content, making it harder to distinguish credible research from automated fabrications.
Taken together, the rise of platforms like Moltbook highlights both the promise and the risks of increasingly autonomous AI systems. As these technologies continue to evolve, society will need to carefully balance innovation with ethical, psychological, and security considerations.
Leave a Reply