Text Embeddings Reveal (Almost) As Much As Text. Embeddings of text—where a text string is converted into a fixed-number length array of floating point numbers—are demonstrably reversible: “a multi-step method that iteratively corrects and re-embeds text is able to recover 92% of 32-token text inputs exactly”.
This means that if you’re using a vector database for embeddings of private data you need to treat those embedding vectors with the same level of protection as the original text.
Recent articles
- Hacking the WiFi-enabled color screen GitHub Universe conference badge - 28th October 2025
- Video: Building a tool to copy-paste share terminal sessions using Claude Code for web - 23rd October 2025
- Dane Stuckey (OpenAI CISO) on prompt injection risks for ChatGPT Atlas - 22nd October 2025