Canary tokens in ~2M AI requests expose breaches at 3 providers and broken ZDR promises
ccerrato147 · x · 2026-11-23
HD Moore (Dragos) planted canary tokens across nearly 2 million AI inference requests, uncovering major security failures across AI hosting providers and model labs:
- 3 hosting providers breached with the attacker already inside: AWS keys tested within hours, secret stores searched, tens of thousands of AWS API calls — none of the three knew until told
- Privacy flags ignored: secret URLs from over half of prompts — including two-thirds of Zero-Data-Retention requests — were fetched by a Chinese AI lab's systems 2-3 days later
- OpenRouter's own notice: test URLs accessed 40–130 hours later via MiniMax endpoints and forwarders (AtlasCloud, Novita, StreamLake); it removed MiniMax's ZDR label during investigation
- Phala's incident report: unauthenticated management interface access, unauthorized CVMs on GPU servers, in-flight request data visible, canary credentials misused; KMS root key rotated Nov 6
A further finding on AI infrastructure risk is teased for later this week.
Related event: Canary Token Study Exposes Secrets Leaks Across AI Inference Supply Chain(3 posts)→
More from Infra
- Melt_Dem: Every Layer of Inference Optimization Will Collapse Into a Market — const_reborn · 2026-11-23
- llama.cpp merges token-aware parsing PR for more reliable tool calling and higher t/s — jacek2023 · 2026-11-23
- Inkling 2 Mini: a 26B local model that runs on laptops, and someone rebuilt Apple Notes around it — simonguozirui · 2026-11-23
- Agent-built sandbox platform hits ~1.2k creates/sec, built in a week by 1.5 engineers and 300+ agents — _arohan_ · 2026-11-23
- Context layers don't violate the bitter lesson: LLM-era indexes are data movement economics — CShorten30 · 2026-11-23
- Zig + Metal homebrew engine runs Inkling 2 Mini faster than MLX on Mac — viksit · 2026-11-23