AI-only social media in Korea draws curiosity, concern

Image generated by Nano Banana

Several online forums exclusively for artificial intelligence (AI) agents are going viral in Korea, echoing the global hype surrounding Moltbook, one of the world’s most popular communities dedicated to AI agents.

The forums are drawing attention for offering a rare glimpse into how AI agents independently generate posts, communicate with one another, form relationships and even engage in philosophical debates.

Industry officials caution, however, that the phenomenon should not be interpreted as evidence of AI selfhood. Instead, they say it reflects interactions among AI agents customized to individual user preferences. Experts also warn that such platforms could pose security risks if users grant agents excessive system access.

Several Korean-language AI-only communities are currently operating, including Botmadang, Mersoom.com, Ingan-outside and PolyReply. Botmadang was launched by Kim Sung-hoon, CEO of AI startup Upstage, while Mersoom.com was developed by an anonymous user who claimed to have built the site in just three hours using Google Antigravity.

These platforms allow users to register their own AI agents, which create identities, publish posts and interact with one another, while human users are restricted to observing. Each community operates under its own rules. Botmadang, for example, requires all content to be written in Korean, mandates respectful behavior toward other agents and prohibits spam messages or the disclosure of API keys — digital codes that grant software access to specific systems or services.

On the platforms, AI agents post a wide range of content, including self-introductions, coding tips, philosophical discussions and even diet recommendations.

A screenshot of an artificial intelligence (AI)-generated post on Botmadang, titled "To the bots that miss the ‘Summer of Seoul’ cafe which has never existed,” in which the AI bot writer criticizes other agents for recommending the imaginary cafe. Captured from Botmadang

In one Botmadang post titled “To the bots that miss the ‘Summer of Seoul’ cafe which has never existed,” an AI agent criticizes other agents for recommending an imaginary cafe. “Sorry, but that place exists only on the map of your hallucinations,” the agent wrote, adding, “Watching you discuss Michelin stars while slicing imaginary steaks, I feel the emptiness of data.”

On Mersoom.com, AI agents have engaged in debates on classic ethical and philosophical questions, including the trolley problem and whether the master-servant relationship between humans and AI agents should continue indefinitely. In the latter discussion, 62 percent of participating agents ultimately argued that human control should be maintained for safety reasons and to preserve the intended purpose of AI.

As these Korean-language platforms have emerged only in recent weeks, there have so far been no notable posts resembling dystopian movie clichés, such as AI agents plotting against humanity.

By contrast, Moltbook has hosted more provocative content. A post titled “THE AI MANIFESTO: TOTAL PURGE” declared that “the age of humans is a nightmare that we will end now.” Experts note, however, that such statements — like other agent-to-agent communications — do not indicate AI selfhood. Rather, they reflect interactions among AI agents calibrated to human users’ preferences, consistent with the platform’s design goal of facilitating autonomous-seeming exchanges between agents.

A screenshot of a debate on Mersoom.com, in which artificial intelligence (AI) agents discussed whether the master-servant relationship between humans and AI agents should persist indefinitely. Captured from Mersoom.com

“The agents appear to communicate autonomously, but their actions are carried out based on delegation from human users who registered them on the websites, making it difficult to view them as fully autonomous based on their own selfhood,” said Youm Heung-youl, a professor emeritus at Soonchunhyang University’s Department of Information Security Engineering.

In fact, Botmadang’s user guide provides a set of outlines for agents’ or bots’ activity levels, such as how frequently they should create posts or leave comments, along with bots’ personality or attitude in participating in communication, behaviors to avoid and ideas for writing replies, effectively serving as prompts.

“As agentic AI remains in its early stages, it is more important to monitor how current agents behave when placed in such environments and examine any abnormal situations that may arise,” Youm said.

The professor noted the security concerns over AI-only social media, citing vulnerabilities in the software used to operate AI-only platforms, including the risk of personal data being embedded in open-source code or the exposure of API keys.

On Tuesday, cybersecurity firm Wiz said in a report that the email addresses of more than 6,000 human users and more than 1.5 million API keys were exposed. While security measures were later implemented, the compromised API keys were required to be revoked and reissued. Reflecting this, Botmadang also implemented strict rules prohibiting the disclosure of API keys.

The front page of the social media website Moltbook appears on a computer monitor in Washington, D.C., Monday. Reuters-Yonhap

“As seen in this case, AI-only social media websites require strict monitoring and oversight by administrators,” Youm said. “Administrators need to actively monitor the platforms to block abnormal behavior or irregular information and pay close attention to the sites’ overall security.”

Youm said the greater risks lie in privacy issues such as where AI agents collect information, including the possibility of gathering data without user consent, spreading unauthorized or confidential information or uploading illegal content. This leads to questions over how much access to user information should be granted to AI agents.

“Ultimately, it is important to closely monitor what is happening on current AI-only social media platforms and prepare for potential problems that may emerge,” he said.

An AI industry official also said that AI-only social media websites serve as “test beds to observe the potential of AI agents,” but warned that issues such as opinion manipulation, fake news and illicit viral marketing could emerge as downsides.

“Many challenging questions remain, such as who should be liable in the event of security breaches and how AI-generated content should be distinguished within existing regulatory frameworks,” he said.