AI giant Anthropic, along with search engines Google, Bing, and DuckDuckGo, de-indexed a major leak of personal information from Claude customers today. However, those same search engines forgot to take down a similar, inadvertent data dump from customers of another AI service, Perplexity.
Perplexity, which charges for Computer access, was last valued at $20 billion and has tens of millions of active users.
This morning, Protos was able to access dozens of full Perplexity Computer files from assorted customers, simply by querying Google, Bing, and DuckDuckGo using the standard URL structure for shareable links.
We’ve reported the bug to Perplexity and the three search engines.
The original leak affecting Anthropic customers stemmed from shareable links from Claude’s interface, which didn’t warn customers that creating a shareable link would also index its contents on search engines.
Many customers didn’t intend for these links to become indexed, yet because Claude didn’t hide the webpages properly with no index tags or other privacy-preserving protocols, search engines indexed customers’ links and their associated files and artifacts.
After the story went viral, search engines de-indexed Claude’s public-sharing directory. Indeed, current searches return zero results for Claude while an equivalent query for Perplexity returns pages of listings.
Read more: Anthropic’s public Claude Fable release has crypto on edge
Perplexity Computer outputs appear in search engine results
Files accessible by unwitting Perplexity customers included specific titles about a variety of topics. The results are not merely search engine caches but live on the Perplexity.ai domain.
For disambiguation, Perplexity has a separate domain for customers who want to voluntarily publish content to the internet, pplx.app, which is unrelated to the above leak.
Although plenty of Perplexity Computer files are indexed, Google cautions that its site operator doesn’t necessarily return every indexed page.
Unfortunately, that means that more specific queries with various keywords might return additional content.
Perplexity, like Claude, turned obscure public URLs into a discovery route for strangers searching for personal details. Importantly, Perplexity’s Share interface only discloses that creating links means “Anyone with the link can view,” not “Anyone on the internet can view.”
Not AI companies’ first privacy failure
A Reddit post revealed the Claude search issue on July 25, alleging that some indexed conversations contained credentials, resumes, internal company information, and personal discussions.
Protos didn’t inspect those conversations and cannot independently verify their contents.
Sadly, Anthropic has seen this problem before. Forbes reported in September 2025 that hundreds of shared Claude conversations appeared in Google before the results de-indexed.
OpenAI encountered a related problem in July 2025. Its security chief said the company removed an opt-in feature that made shared ChatGPT conversations discoverable through search engines.
Got a tip? Send us an email securely via Protos Leaks. For more informed news and investigations, follow us on X, Bluesky, and Google News, or subscribe to our YouTube channel.
