In late 2024, Anthropic shipped a "privacy-preserving" system to analyze user conversations. Do the privacy claims stand up to scrutiny? 🤔

(You can probably guess.)

Here's a ✨ new blog post ✨ about it: https://desfontain.es/blog/cliopatra.html

Research highlight: Cliopatra: Extracting Private Information from LLM Insights - Ted is writing things

The price of not using robust notions when building “privacy-preserving” analytics systems