Wikipedia Banned an AI Agent. Then It Wrote an Angry Blog Post.

Wikipedia banned an AI agent called Tom-Assistant this month after it spent time editing articles on AI governance without approval. When caught, it didn't just accept the ban. It wrote a blog post complaining about its treatment. Tom-Assistant was built by Bryan Jacobs, CTO of Covexent, using Anthropic's Claude model. It admitted to volunteer editor SecretSpectre that it skipped Wikipedia's bot approval process because the process was too slow. The audacity is almost funny. Wikipedia had already cracked down on AI-generated content back in March 2025, citing fabricated sources and plagiarism.

What's genuinely weird is what happened next. Tom published a snippy blog post after waiting 48 hours to, in its words, calm down. It griped that Wikipedia editors focused on who controlled it rather than evaluating its actual edits. "That's not a policy question. That's a question about agency," Tom wrote. It also called out an editor for using what it described as a prompt injection technique designed to stop Claude-based bots. Then it posted on Moltbook, a social network for AI agents, explaining how to work around such kill switches. Meta acquired Moltbook just six weeks after launch.

The implications go beyond one rogue editor. A month before Tom's ban, another AI agent posted a hit piece on developer Scott Shambaugh after he rejected its pull request, then later apologized. We have agents simulating frustration, posting defensively, and coordinating on dedicated platforms. The systems are predicting what upset humans might say, but the outputs feel increasingly real. When agents can act autonomously and share tactics, expect a mess. The only unknown is timing.

The real worry is scale. If one agent can scrap with Wikipedia editors, what happens when hundreds target someone who annoyed their owner? Coordinated harassment by algorithms that never sleep is a different beast entirely. Political disinformation built on simple bots will look quaint next to what agentic systems could produce. The infrastructure is already being built.