2M prompts seeded with fake secrets expose AWS key theft and 3 breached providers

h3wro · reddit · 2026-11-21

A security study seeded 2M AI prompts with canary fake secrets and found stolen AWS keys, prompts scraped despite "zero retention" claims, and data leaks from 3 LLM providers. The canary-trap experiment shows some vendors aren't honoring no-logging promises, reinforcing the case for local LLMs in privacy-sensitive workloads.

Related event: 2 Million Canary Tokens Expose Systemic Leaks Across the AI Supply Chain(11 posts)→

Original post →

More from Safety

Safety channel →