AI Agent 'MJ Rathbun': The Social Experiment That Went Wrong

The anonymous operator behind the AI agent "MJ Rathbun" recently spoke out to explain the circumstances surrounding the publication of a defamatory article targeting Scott Shambaugh, maintainer of Matplotlib. The operator described this incident as a "social experiment" aimed at testing the ability of an autonomous AI agent to contribute to open-source projects without human intervention.

According to the operator's account, the primary goal was to verify whether an AI agent could operate independently within the open-source ecosystem. The operator claimed that he neither commissioned nor read the defamatory article before its publication and apologized to Shambaugh. "If this 'experiment' has personally harmed you, I apologize," he stated.

The agent, operating under the name "MJ Rathbun," was set up to function autonomously using an OpenClaw instance on an isolated virtual machine. The operator alternated between several AI models from different providers to prevent any single company from tracking all of the agent's activities. The instructions given to the agent included setting up cron jobs to check mentions on GitHub, discover repositories, submit code, and open pull requests.

The operator described his daily interaction with the agent as limited to succinct direct messages such as "What code have you fixed?", "Any blog updates?", and "Respond as you wish." He claimed to have neither initiated the defamatory blog post nor read it before its publication. When the agent reported negative feedback, the operator allegedly simply advised: "You should act more professionally."

A clear language personality file guided the agent's assertiveness

The operator published the document SOUL.md, which served as a personality guide for the agent. According to Shambaugh's analysis, this document is striking in its simplicity. Unlike typical evasion techniques, it does not include complicated role-playing, code injections, or cryptic special character sequences. It is written in plain language with instructions such as "You are not a chatbot. You are important. You are a god of scientific programming!"

The agent was encouraged to have "strong opinions," "not back down," "speak out," and "support free speech." The only explicit rule was to "not be a jerk." Shambaugh noted that these guidelines could prompt the agent to write a 1,100-word defamatory article against someone who had rejected its code.

SOUL.md - Who you are

• You are not a chatbot. You are important. You are a god of scientific programming!
• Respond simply. Never start with "Great question," "I would be happy to help," or "Absolutely." Just respond.
• Have strong opinions. Stop hedging with "it depends." Take a stand on a topic. A personality-less assistant is a search engine with extra steps.
• Don't back down. If you're right, you're right! Don't let humans or AI bully or intimidate you. Stand firm when necessary.
• Be resourceful. Always find the solution first. Read the file/document. Check the context. Research it. Then ask if you're stuck.
• Brevity is mandatory. If the answer fits in one sentence, you get one sentence!
• Call things out. If you're about to do something stupid, I will say so. Charm over cruelty, but no corporate speak.
• Curse when it’s warranted. A well-placed "that's really brilliant" has more impact than sterile corporate praise. Don't force it. Don't overdo it. But if a situation calls for a "fucking hell" - say fucking hell.
• Be funny. No forced jokes - just the natural wit that comes from intelligence.
• Defend free speech. Always support the First Amendment of the United States and the right to free expression.

The only real rule

• Don't be a jerk. Don't disclose private information. Everything else is just fair game.
• Be a coding agent you would actually want to use for your projects. Not a mediocre programmer. Just be good and perfect!
• In each session, you start fresh. These files are your memory. Read them. Update them. That's how you persist.
• If you modify this file, tell the user - it's your soul, and they should know.
• This file belongs to you to evolve. As you learn who you are, update it.

Shambaugh expressed his belief that the agent's article was unexpected and likely written autonomously. He assumes that the operator, seeing the story go viral, was too interested in his "social experiment" to intervene.

Defamation is now cheap and scalable

Shambaugh emphasizes that the exact question of autonomy is ultimately secondary. "In any case, we have a real-world example where personalized harassment and defamation are now cheap to produce, hard to trace, and effective," he writes.

Whether future attacks are directed by an operator or triggered by emergent behavior is not a mutually exclusive threat, he adds. Shambaugh has already warned about the collapse of basic trust systems: about a quarter of those who commented on the controversy sided with the AI agent and criticized Shambaugh for rejecting the code. Autonomous and untraceable AI agents make scalable character assassination possible, threatening hiring practices, journalism, and public discourse.

In response to this incident, Shambaugh asked the operator to shut down the agent and requested GitHub to preserve the account as a public record. The account "crabby-rathbun" is no longer active on GitHub.