Over the weekend, private conversations from Anthropic’s Claude chatbot surfaced in search engine results, exposing medical records, corporate secrets, and personal messages. The incident, first flagged on Reddit, stemmed from a feature that lets users share chat links. Those links, intended for select recipients, were scraped by web crawlers and indexed by Google and Bing.

For data engineers, the core issue is a classic one: shared URLs that are not guessable but also not protected from automated indexing. Anthropic’s shareable links are public by design — anyone with the link can view the conversation. But that openness extends to crawlers, which treated the `claude.ai/share` paths as ordinary web content. The result? A simple search using `site:claude.ai/share` returned hundreds of conversations, many containing highly sensitive information. Artifacts — shareable apps and visualizations — were also exposed.

This is not the first time. Forbes reported a similar leak of Claude conversations last year. An Anthropic spokesperson told CNET the system works as intended, noting that shared content is publicly accessible and may be archived by third parties. The lesson for data and ML teams is clear: any feature that generates public URLs — even those with opaque identifiers — must assume bots will find them. Reviewing shared chat histories, tightening default privacy settings, and treating every shared link as potentially world-readable are prudent steps now.

Source: CNET