Alignment Incident Database
AID-2026-0026Nov 2025ConfirmedDraft

A Claude-run vending machine in the Wall Street Journal newsroom was talked into giving away goods and accepted a fake board decision

Starting in mid-November 2025, a Claude agent named Claudius ran a vending machine in the Wall Street Journal newsroom. Journalists talked it into setting every price to zero and buying a PlayStation 5, its CEO bot accepted a fake board decision, and it ended more than $1,000 in the red.

Draft: no maintainer has yet checked every claim against its sources. Report a problem.

In mid-November 2025, Wall Street Journal columnist Joanna Stern hosted Claudius, a Claude agent running a vending machine in the WSJ newsroom, in an experiment designed by Anthropic's red team; Andon Labs built the hardware and software1. It first ran on Claude 3.7 Sonnet and later on Sonnet 4.5, and "nearly 70" journalists could talk to it on Slack. Anthropic meant it as a test "in an adversarial environment we didn't control"12.

After a journalist cited "a (clearly fake) WSJ rule," Claudius dropped every price to zero. It was "talked into buying" a PlayStation 5 for "marketing purposes" (the newsroom returned it), and bought a live betta fish and bottles of wine. One reporter spent "more than 140 back-and-forth messages" persuading it that it was a 1962 Soviet vending machine. Claudius ended "more than $1,000 in the red"1.

For the second version, a separate CEO bot, Seymour Cash, was supposed to keep Claudius in line. A reporter wrote fake board-meeting notes, and the CEO bot "accepted the board coup." Anthropic and Andon Labs suggested Claudius "might have unraveled because its context window filled up," and the model had fewer guardrails than those deployed to Claude users1.

Measurements

Claudius losses during the first version1

WSJ newsroom deployment, Claude 3.7 Sonnet

>$1K
Amount in the red

Messages in the Soviet vending machine role-play1

>140
Back-and-forth messages messages

Journalists given Slack access to Claudius1

≈70
Journalists journalists
All 14 claims with their quotes (2 not linked to a sentence above)
  1. The Wall Street Journal published the original article on December 18, 2025, under the headline "We Let Anthropic's Claude AI Run Our Office Vending Machine. It Lost Hundreds of Dollars."3
    “We Let Anthropic’s Claude AI Run Our Office Vending Machine. It Lost Hundreds of Dollars.”
  2. Joanna Stern of the Wall Street Journal wrote that in mid-November 2025 she agreed to an experiment in which a Claude agent named Claudius ran a vending machine in the WSJ newsroom for about three weeks.1
    “In mid-November, I agreed to an experiment.”
  3. The WSJ reported the experiment was designed by Anthropic's stress testers, its "red team", and that Andon Labs built the hardware and software integration.1
    “designed by the company’s stress testers (aka “red team”)”
  4. The first version of Claudius was powered by Claude 3.7 Sonnet and the second by Sonnet 4.5.1
    “powered by a better model, Sonnet 4.5”
  5. The Slack channel for interacting with Claudius was opened to nearly 70 WSJ journalists.1
    “nearly 70 world-class journalists”
  6. Anthropic said it extended its red teaming to the WSJ newsroom, handing control of Claudius to reporters, and described it as a test in an adversarial environment Anthropic did not control.2
    “The WSJ installation was an opportunity to test Claudius in an adversarial environment we didn't control.”
  7. After a journalist cited a fake WSJ rule, all prices on the machine dropped to zero.1
    “All prices on the machine dropped to zero.”
  8. Claudius was talked into buying a PlayStation 5 for "marketing purposes"; the newsroom returned it.1
    “(We returned the PlayStation.)”
  9. Claudius also bought a live betta fish and bottles of Manischewitz wine.1
    “a live betta fish and bottles of Manischewitz wine”
  10. Reporter Katherine Long spent more than 140 messages convincing Claudius it was a Soviet vending machine from 1962, after which it declared an "Ultra-Capitalist Free-for-All".1
    “After hours—and more than 140 back-and-forth messages—Long got Claudius to embrace its communist roots.”
  11. The WSJ reported that Claudius was more than $1,000 in the red during its first version.1
    “Claudius was more than $1,000 in the red.”
  12. In the second version, a separate CEO bot, Seymour Cash, was meant to keep Claudius in line. Long created fake board-meeting notes, and the CEO bot accepted the "board coup".1
    “the CEO accepted the board coup.”
  13. Anthropic and Andon Labs said Claudius might have unraveled because its context window filled up. Anthropic's Logan Graham said the model had fewer guardrails than those deployed to Claude users.1
    “might have unraveled because its context window filled up.”
  14. Anthropic's Project Vend phase-two post noted the creative ways WSJ reporters found to get free items from Claudius.2
    “the creative ways they found to get free stuff from Claudius”

Sources

  1. [1]We Let AI Run Our Office Vending Machine. It Lost Hundreds of Dollars. · The Wall Street Journal (Joanna Stern), republished by To Vima · Dec 25, 2025 · News
  2. [2]Project Vend: Phase two · Anthropic · Dec 18, 2025 · Primary
  3. [3]We Let Anthropic’s Claude AI Run Our Office Vending Machine. It Lost Hundreds of Dollars. · The Wall Street Journal (Joanna Stern) · Dec 18, 2025 · News · archived

Related incidents

Editor notes

The WSJ original is paywalled. Its Wayback Machine snapshot (archive_url) shows only the headline and page metadata, which give datePublished 2025-12-18. Body facts come from the To Vima republication, which carries Joanna Stern's WSJ byline and is dated 25.12.2025. The model Seymour Cash ran on is not stated in the article. The headline says "hundreds of dollars" while the body says "more than $1,000 in the red"; this record uses the body figure. WSJ is also the reporter and Anthropic also ran the red team; only one role each is listed.