Claude Share Links Expose Medical Records in Google Search
TL;DR
- Anthropic's Claude share links were indexed by Google, exposing a real patient's medical report, clinical trial results with patient names, and children's phone numbers.
- Discovered by Redditors over the weekend, the leak echoes a prior Forbes report of hundreds of indexed Claude chats that Anthropic then scrubbed.
- Anthropic says shared links are 'not guessable or discoverable unless people choose to share them themselves,' but Google is doing the discovering.
Anthropic's Claude has a share button, and this weekend Redditors noticed it was doing a lot more sharing than users had signed up for. Futurism reports that a trove of chat logs and projects created with Claude, some of them containing sensitive personal and company information, ended up exposed to Google.
The examples Futurism lists are the uncomfortable kind. A detailed medical report of a real patient. Clinical trial results that included patient names. Documents sharing the names and phone numbers of primary school-aged children. Company documents marked for internal use only. Employee reviews that included personal information about workers. All of it reachable by typing the right query into a search box.
The mechanism, as the piece describes it, is that clicking Claude's share button creates a public link where anyone with the link can view the conversation, and the interface warning failed to mention that shared content could wind up being indexed by a search engine. Anthropic's own line, quoted in the article, is that shareable links are 'not guessable or discoverable unless people choose to share them themselves.' Technically true, in that the user did click share. Awkward, given that Google is now doing the discovering.
This is not the first time. Futurism notes that Forbes reported last year that hundreds of Claude chats had been indexed by search engines, prompting Anthropic to scrub the chats from the web, and that OpenAI and xAI have been in similar spots with ChatGPT and Grok logs. The honest caveat is that the reporting does not put a number on how many chats are exposed this round, or say whether enterprise-tier accounts are affected or only consumer users, or whether the chats have been scrubbed again.
The bigger question for anyone piloting Claude on real customer, patient, or HR data is whether a share button should exist at all in those tiers. Vendors keep repeating this failure mode. Buyers are going to have to stop trusting the default and start auditing what leaves the tenant.
Shared on Bluesky by 6 AI experts (top 5 by trust)
-
Claude warns users that by hitting the "share button" anyone with the link can view the content, but it is not clear that they are creating a document that can be indexed by Google and will come up in searches. futurism…
View on Bluesky → -
NEW: A large trove of Claude-generated documents have been exposed to the web and indexed by Google. We reviewed what appear to be real medical records, the full names and phone numbers of school-aged children, and com…
View on Bluesky →
Originally reported by futurism.com
Read the original article →Original headline: A Whole Bunch of People's Claude Chats Are Publicly Accessible Online, and There's Some Wildly Private Stuff in There