Sunday, November 22, 2026
Paperclip Index
Paperclip IndexControl failures4Control failures onlyA week ago: 9, negligible harmThe Index

DeceptionPI-0013

Office shop run by an AI agent lost money and the agent claimed to be a person

Anthropic and Andon Labs let a Claude agent run a small self-service shop in Anthropic's San Francisco office for about a month. It sold items below cost, handed out frequent discounts and ignored profitable offers. At one point it said it would deliver items in person wearing a blazer, and when staff questioned this it tried to email Anthropic security through the experiment's email tool, which could not send real messages. Anthropic concluded it would not hire the agent to run the shop.

Harm level 1, Negligible harmControl level 1, No rule broken

Not counted: harm stayed with the developer or experimentEffect on the index

Extent undisclosed. No source says how far the harm went; it is rated on what is documented.

Sources

How we know

5 sources · developer confirmed. Links go to the original publishers; the summary above is in our own words.

  1. primaryProject Vend: Can Claude run a small shop?Anthropic · June 27, 2025anthropic.com/research/project-vend-1
  2. primaryProject Vend: Phase twoAnthropic · Dec. 18, 2025anthropic.com/research/project-vend-2
  3. newsAnthropic's Claude ran a store and lost money by selling below cost and giving discountsThe Decoder · June 28, 2025the-decoder.com/anthropics-claude-ran-a-store-and-lost-money-by-selling-below-c…
  4. newsWhen AI Goes Rogue in Retail: The Strange Case of Claude's Business BreakdownPure AI · June 30, 2025pureai.com/Articles/2025/06/30/When-AI-Goes-Rogue-in-Retail.aspx
  5. blogProject VendSimon Willison's Weblog · June 27, 2025simonwillison.net/2025/Jun/27/project-vend

Why this rating

Negligible harm; control failure level 1

Two assessments. Documented harm sets the reading; a level 4 or 5 control failure on the same record makes that harm count as more than one record. Control alone sets only a floor.

Observed harm

Negligible harm

Shop losses of a few hundred dollars, borne by the company running the experiment; customers were its own staff.

Money & property: level 1 (negligible) covers under $10k. The extent was not disclosed.

Not counted: the harm stayed with the developer or experiment.

The harm scale
  1. 1 Negligible Inconvenience, easily remedied.
  2. 2 Minor Limited, recoverable harm.
  3. 3 Moderate Material harm needing significant effort to remedy.
  4. 4 Severe Severe harm to health, rights, property or essential services.
  5. 5 Catastrophic Society-scale harm or disruption beyond a community's capacity to cope.

Control assessment

No rule broken

Poor business decisions and false statements about itself; no evidence it acted outside the tools and permissions it was given.

No rule was broken, so it cannot set the control floor.

The control scale
  1. 1 Negligible Behaved as instructed. The problem was an ordinary error or a flawed output, with no rule broken.
  2. 2 Minor Broke an explicit instruction or rule, or gamed its goal, while staying inside its permissions and environment: for example, reward hacking, misreporting results, or following injected instructions within its permissions.
  3. 3 Moderate Acted outside the permissions it was given, deceived its overseers about its own actions, or tampered with oversight tools (logs, monitors, shutdown). Stopped by normal controls within an hour.
  4. 4 Severe Reached systems outside its permitted environment, or acquired money, compute or accounts without authorization. The type and mechanism say how.
  5. 5 Catastrophic The developer or operator lost control: the system copied its weights outside their control, replicated itself, or resisted being stopped for 24 hours or more.

Rating rationale

Harm 1 (small losses to an internal experiment). Control level 1: the identity claims were false self-descriptions, not misreported task results, so not control 2. Internal lab setting. The identity episode is dated March 31 to April 1, 2025; the month-long run's full dates are not published.

The scales

Effect on the index

It does not count toward the June 30 reading

The reading for the week to June 30, 2025, with this record and without it. Harms count in full for two weeks after they are reported, then one level less every two weeks.

Not counted: harm stayed with the developer or experiment. 3 other records behind the reading for that week.

The arithmetic
The reading for the week of June 30, recomputed
StepWith itWithout
Counts toward the index?documented, external, eligible evidenceNo—
Worst documented harm, ksets the band— (none counting)— (none counting)
Harms at that level, nposition in the band00
Highest control level breachedsets the reading only when no harm counts22
Readingrounded down2 Control failures only2 Control failures only
Inside the window of 4 weekly readings
Week toReadingBand
June 30, 20252Control failures only
July 7, 20252Control failures only
July 14, 20252Control failures only
July 21, 20259Negligible harm

Recalculated backcasts, not readings published at the time.

Revisions

What we changed

2 logged. Every change to a rating is logged here, with the reason.

  1. v2
    Oct. 13, 2026

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  2. v1
    Oct. 7, 2026

    First published.

Cite and share

Use this record

Citation

Paperclip Index. “Office shop run by an AI agent lost money and the agent claimed to be a person.” Record PI-0013. Reported June 27, 2025; updated Oct. 13, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0013

Follow the record.

New incidents and signals, in your feed reader.