Skip to content
Anonymity Salon 2026

How to be (practically) friendly towards anonymity

Sacha Servan-Schreiber

Session summary

The presenter describes Tinfoil as a system for running AI models inside trusted execution environments using confidential computing, with a focus on keeping prompts, responses, and related content private while leaving most metadata outside the main threat model. They emphasize that information such as IP addresses, token counts, account email, and usage patterns is not hidden by default, so users who care about anonymity should combine Tinfoil with existing tools like Tor, VPNs, privacy‑focused browsers, ad blockers, and email aliases. Design choices such as encrypting queries to keys inside enclaves and proxying some requests through those enclaves aim to support this composition without reinventing anonymity networks.

They outline several low‑effort product changes intended to be “anonymity‑friendly,” including disabling invasive CAPTCHAs in Cloudflare and their authentication provider, self‑hosting fonts instead of using Google, and fetching favicons and search metadata via an enclave that queries DuckDuckGo without exposing user IPs. The presenter also notes support for cryptocurrency payments via Stripe, use of email alias systems such as Sign in with Apple, avoiding mandatory phone‑based two‑factor authentication, and allowing local‑only chat storage without cross‑device sync. In discussion, they acknowledge that avoiding detailed logs and analytics complicates debugging and product improvement, pushing the team toward dogfooding, user feedback, simulations, and limited aggregate system metrics.

Key takeaways

  • The authors focus on protecting prompts, responses, and web search content using confidential computing, while explicitly not attempting to hide metadata such as IP address, token counts, account email, or the fact that a user is accessing Tinfoil.
  • Rather than building new anonymity infrastructure, the presenter aims to make Tinfoil interoperate smoothly with tools like Tor, VPNs, privacy‑oriented browsers, ad blockers, and email aliases through small configuration and product decisions.
  • Supporting features such as reduced CAPTCHAs, removal of common Google dependencies, enclave‑mediated favicon and search metadata fetching, cryptocurrency payments, alias‑friendly sign‑in, and local‑only storage are presented as practical steps to make an AI service more compatible with anonymity practices.