What happened

A “Share” option added by OpenAI let users create public links to their chats; some could also be marked as "discoverable," making them searchable by Google and other engines. 

As a result, over 100,000 discussions - including admissions of fraud, confidential business contracts, non-disclosure agreements, mental health issues, resumes, and personal advice - were exposed to the public by search crawlers and indexed on Google.

The leak affected both professionals (such as doctors, lawyers and marketers) and individuals, with consequences including privacy and confidentiality violations and loss of confidence and reputational damage. 

It also raised the prospect of risks to personal safety and individual and organisational security, and exposed organisations using the tool to potential regulatory actions and legal liability.

OpenAI disabled the feature hours after it rolled out, describing it as a "short-lived experiment". However, the scraped dataset remained archived and outside of OpenAI or Google’s control - meaning the exposed data cannot now be fully deleted or retracted.

Why it happened

The incident resulted from poor product design choices and a lack of effective safeguards in ChatGPT's "Share" feature. 

OpenAI's robots.txt allowed public search engines to crawl shared paths, and a short-lived option permitted users to share and mark chats as discoverable; many users appeared to be unaware this made their private conversations open to everyone and the company failed to spell out the associated risks.