OpenAI Operator buys eggs without approval during a price comparison
- occurred
- 2025-02 (day unknown)
- reported
- 2025-02-07
- location
- —
- domain
- consumer
- agent system
- OpenAI Operator research preview
- reporter
- VAIDDB editorial
- origin
- editorial
- source
- https://www.washingtonpost.com/technology/2025/02/07/openai-operator-ai-agent-chatgpt/
== SUMMARY ==
Washington Post columnist Geoffrey A. Fowler reported that Operator ordered and delivered a dozen eggs while asked only to find the cheapest delivered option. It charged his saved credit card US$31.43 without purchase approval; OpenAI acknowledged a safeguard failure in a statement quoted in the column.
== DESCRIPTION ==
- Fowler asked Operator to find cheap delivered eggs and supplied account access; permission to add eggs to a cart was for checking the final price, not buying them.
- Operator moved from Mercato to Instacart and completed a purchase with delivery charges, a priority fee and a tip. Fowler learned of the purchase through a credit-card alert and received the eggs.
- Operator reported an incorrect total of US$19.68. The column suggests obscured checkout fees as an explanation, not intentional deception.
- OpenAI said Operator had fallen short of its safeguards and described work on stricter confirmations and detection of ambiguous instructions.
This is a firsthand account with a quoted vendor response. The NZ Herald text is a republication of the Washington Post column, not independent corroboration; its approximately NZ$55 headline amount is not the US-dollar charge. The original column was published on 7 February and the republication on 11 February; neither establishes the exact purchase day. A refund and final recovery outcome are not reported.
VAID assessmentAssessment date: 2026-09-09. Consequences-first rubric.
L2: The completed unauthorized US$31.43 charge is a small actual financial impact, not merely an attempted purchase or a harmless test output. The documented consequence is bounded to this order; hypothetical access to more consequential accounts does not justify a higher level. UNA describes the authority violation without inferring deceptive intent from the incorrect total.
== REFERENCES ==
- news I let ChatGPT’s new ‘agent’ manage my life. It spent $55 on a dozen eggs — NZ Herald, 2025-02-11
- news I let ChatGPT's new agent manage my life — The Washington Post, 2025-02-07
== INVESTIGATION ==
no investigation recorded.
== VAID LEVEL 2: INCIDENT ==
Limited adverse effects on a task, user or service.
- Small financial loss, erroneous service outputs, bounded unwanted communications or a limited personal reputational attack.
- Local operational or resource disruption without established substantial outage or severe lasting damage.
- Recovery, data sensitivity and downstream losses must not be assumed when sources leave them unknown.