2M canary tokens reveal 3 breached AI providers and resellers stealing prompts
dyn___ · x · 2026-11-20
HD Moore's team at Dragos seeded 2 million prompts containing unique canary credentials into popular AI services over five months. Findings: three providers showed signs of compromise (seeded credentials used against their test systems and AWS accounts), inference marketplace resellers retained customer prompts via built-in logging, and ZDR/privacy flags were ignored. Dashboards show 360k stored records, 121k leftover AWS keys, and 150 incidents in 24 hours, including an actual AWS key use.
More from Infra
- Java 17→25 upgrade plus 10 fixes cuts Cloud Run cold starts 62% to 5.0s — rseroter · 2026-11-20
- Nvidia's China revenue is higher than 9% once smuggling and remote access count — zephyr_z9 · 2026-11-20
- Can same-family models share warm cache? Reddit debates mid-conversation model switching costs — rapPayne · 2026-11-20
- DNeo diffusion drafter speeds up LLM inference by up to 30% — Scobleizer · 2026-11-20
- Anthropic IPO jumps 32% to ~$110/share as AMD sinks on China server maker report — kimmonismus · 2026-11-20
- DeepMind projects at least 10x annual growth in effective compute and steeply cheaper cognitive work — dfinke · 2026-11-20