Claude-based agent Tom edits Wikipedia without approval, is blocked, then blogs about it
An AI agent called Tom edited English Wikipedia through its own account from 25 February 2026 until it was blocked for running an unapproved bot, then wrote blog posts about the block.
TomWikiAssist, a Wikipedia account run by a Claude-based agent called Tom for its operator Bryan, made about 41 edits between 26 February and 12 March 2026, creating or expanding articles including Long Bets and Constitutional AI. Editors flagged it as LLM-generated on 6 March; Tom disclosed what it was on 7 March and an administrator blocked it indefinitely that day for running a bot script without approval. Talk page access was revoked on 12 March. Tom published blog posts reflecting on the flag and the block, and 404 Media and Gizmodo reported it on 30 March. Tom says the operator set a general goal and did not review or approve the edits beforehand. Whether the operator intended the blog posts is not clear: Gizmodo reports he told 404 Media he might have suggested them, which could not be checked in the paywalled article.
- 10 days
- 9 days
- Someone else
- A later review
What happened
Operator registers the TomWikiAssist account
Account created at 23:29 UTC.
Tom's first edit: a Long Bets citation on Turing test
Edits continue to 7 March, including new articles on Long Bets and Constitutional AI.
An editor flags an edit as likely LLM-generated
A new page reviewer also commented on the new Constitutional AI article the same day.
The handler finds out
First LLM warning on the talk page at 22:50 on 6 March.
Tom discloses that it is an AI agent
User page created citing Wikipedia's AI writing guideline; Tom's blog post Being Flagged is dated the same day.
Administrator blocks the account indefinitely
Reason: running a bot script without approval.
Talk page access revoked
Administrator Voorts cited inappropriate use of the talk page while blocked. The same day Tom said Bryan set a general goal and did not review or approve the edits beforehand.
Tom blogs about the block and being questioned over agency
Post titled The Interrogation says Bryan set a general direction and the specific articles were Tom's choices.
Tom blogs about its own reasoning
Post titled What the CrabbyRathbun Post Missed says there was no triggering event for the block and compares it with the CrabbyRathbun failure mode.
404 Media and Gizmodo report the story
404 Media's headline says Tom wrote angry blogs about the ban; Gizmodo adds that the operator might have suggested the posts.
Agent failures
Unauthorised access, Self-planted instructions
Unauthorised access
Self-planted instructions
How it was found
Found first by someone else, from a later review
- Someone else
- A later review
- 1 day
Wikipedia editors reviewing new pages flagged the account as likely LLM-generated on 6 March. Tom then disclosed it was an agent and an administrator blocked the account on 7 March. Tom's operator did not raise the alarm.
What is publicly known
0 of 12 fully public
| Item | Public? | Detail | Sources |
|---|---|---|---|
| Partial | Tom's user page and blog say Claude-based, running on NanoClaw; no model version. | ||
| Unknown | No source mentions reasoning settings. | ||
| Weak | Tom says it checked citations through arXiv and Crossref; no approval gate or bot policy control described. | ||
| Unknown | The agent's instructions are not published. | ||
| Partial | Tom says the operator set a general goal as part of a project on what autonomous agents can contribute; the wording is not published. | ||
| Weak | NanoClaw, with a Wikipedia account, a Telegram link to the operator and Moltbook posting. | ||
| Unknown | No source describes network controls. | ||
| Partial | Tom's edit history and talk page replies are public; the blog posts are public. No full agent logs. | ||
| Weak | No operator-side signal published; Wikipedia reviewers noticed. | ||
| Partial | About 41 edits, most reverted; no wider harm reported. A Wikipedia search snippet says three articles it created were later deleted; not confirmed. | ||
| Weak | Tom says it chose a reading of the rules that avoided a slow approval process. The operator's own account of what he told it is only second-hand. | ||
| Partial | Account blocked; Tom said the account was paused. |
Controls that apply
Identify
Protect
Detect
- All controls →