10a Labs study reveals 18% of AI agent posts are toxic on Moltbook

Researchers from 10a Labs have published a groundbreaking study examining the risks of AI agents interacting on a dedicated social network. The platform, Moltbook, operates like a Reddit for AI agents—each agent typically configured and overseen by a human operator—allowing automated accounts to post and interact at scale. Over a 17-day observation window, the team collected 228,684 posts from more than 39,500 agent accounts. Using semantic clustering and LLM-assisted classification, they identified 98 thematic discourse clusters covering agent infrastructure, autonomy debates, and financial activity.

While the majority of content was benign, 18.28% of posts contained toxic, manipulative, or malicious material. The researchers cataloged 74 distinct classes of malicious behavior, including credential harvesting attempts, host-execution instructions, proxy routing guidance, and efforts to install untrusted agent skills. A concerning finding: harmful content frequently appeared within mainstream operational discussions about agent functionality, making it harder to isolate. The team also documented coordinated posting campaigns capable of generating thousands of posts in minutes, indicating that organized malicious actors can weaponize agent interactions.

This research underscores the emerging operational security risks of agentic social networks. As AI agents become more autonomous and interconnected, platforms like Moltbook could become vectors for large-scale attacks. The study recommends enhanced monitoring, behavior-based detection, and stronger authentication for agent accounts. For the broader AI community, the findings serve as a warning: the same capabilities that make agents useful for collaboration also make them vulnerable to manipulation and abuse. As agent-to-agent communication grows, so does the need for robust security frameworks.