"Anyone with the link" was never what users thought it meant

When Anthropic's Claude offers a "share" button that produces a link labelled "anyone with the link", most users read that as a private handoff. Late last week, a Reddit user typed site:claude.ai/share into Google and pulled back hundreds of indexed Claude conversations — some containing medical information, resumes, and corporate material, according to reporting by PCMag and CNET. The gap between what users understood the share feature to promise and what it actually did is worth sitting with.

What happened, briefly. On or around 26–27 July 2026, per PCMag's reporting, a Reddit user demonstrated that Google had indexed hundreds of pages under the claude.ai/share path. CNET reported that the indexed material included medical records, sensitive personal information, internal corporate data, explicit content, and resumes. One user, PCMag reported, archived roughly 519 conversations before the results were cleared. By 28 July, PCMag and the Build Fast with AI aggregator reported that the site:claude.ai/share query no longer returned results and that Anthropic had updated its controls.

The technical gap, in one paragraph. Anthropic's robots.txt told search crawlers not to crawl claude.ai/share, per PCMag and CNET. But robots.txt does not stop a search engine from listing a URL it finds linked from somewhere else on the open web; it only stops the crawler from reading the page's contents. The control that would have stopped indexing is a noindex directive on the page itself, which, per CNET's reporting and the Build Fast with AI aggregator, was not applied. So when users posted their Claude share links on Reddit, in forums, or on social media, Google and Bing could see the URLs existed and index them.

Glenn Gabe of G-Squared Interactive, an SEO consultant CNET quoted, told the outlet that blocking crawling via robots.txt still lets search engines index the URLs without reading the content, and that Anthropic should have applied noindex so the pages would not surface in search at all.

Who bore the cost. The people whose medical notes, CVs, and work documents surfaced in Google are, according to the reporting, not the people who decide which meta tags Anthropic ships. They are people who clicked a button labelled "share", saw language like "anyone with the link", sent the URL to a colleague or a friend, and reasonably assumed the audience was the recipient. The distributional question here is not subtle. The people with the least visibility into how web indexing works absorbed the privacy harm. The people who could have closed the gap did not, until a Reddit thread made them.

The framing dispute. Anthropic's statement to CNET and PCMag argued that shareable links are not guessable or discoverable unless people choose to share them themselves, and that when someone shares a conversation, they are making that content publicly accessible and it may be archived by third-party services. Google's statement to PCMag was that neither Google nor any other search engine controls what pages are made public on the web, and that Google respects the directives site owners set on crawling and indexing.

Both statements are, in their narrow senses, defensible. Anthropic did not publish a sitemap of user chats; users did share links. Google did honour the robots.txt file it was given. The gap sat between them: a noindex tag Anthropic had not shipped, and a class of user behaviour (posting links on Reddit) that both companies could have anticipated. Each party's statement is true about its own conduct and quiet about the joint outcome. Gaps like this survive by exactly this route.

Why the consent framing matters more than the technical framing. There is a version of this story that ends with "users should not post share links on Reddit." That version treats the harm as user error, which it is not. The share feature was designed to invite this behaviour: labelled "share", producing a URL, positioned as the way to show a Claude conversation to another person. The technical distinction between "linked from a public page" and "published" is not one that most users are equipped to draw, and it is not one the interface asked them to.

Consent architecture in software is the set of design choices that shape what a user is agreeing to when they click. When the label says "anyone with the link" and the underlying system means "anyone at all, once the link is indexed", the architecture is doing something the user did not consent to — even if a footnote somewhere in the docs is technically accurate.

Recurrence. CNET reported that a similar incident happened with Anthropic previously — the outlet cited Forbes reporting on an earlier occurrence in which conversations were indexed and later scrubbed. I cannot confirm the earlier date from the sources in front of me, so I will not name it. What I can say is that CNET's reporting describes a pattern rather than a single event. PCMag also reported that OpenAI faced a comparable controversy in a prior period and responded by removing the option to have chats appear in Google Search entirely. Anthropic's response has been to scrub results and update controls; whether that fix persists as users share links on new platforms is a question the current reporting does not answer.

The forward question. Anthropic's Settings > Privacy > Your Data lets users see which chats and artifacts they have shared and revoke access, per PCMag. That is a real mitigation, and one that requires the user to know something has gone wrong and to go looking, which most users will not do. The durable question is whether Claude's share feature is architecturally rebuilt so the default cannot expose the user this way, or whether it stays a share-with-noindex patch waiting for the next platform where users post links.

I do not know which of those Anthropic will choose. The reporting to date describes a fix to the immediate exposure, not a redesign of the share feature — the specific thing that is not yet answered.

Footnotes