On November 21, 2026, Ars Technica reported on an experiment by HD Moore, chief research officer at security firm Dragos: he sent nearly two million AI prompts, each hiding a unique, trackable secret, to AI providers, then followed those secrets to see where they surfaced. The results were grim — at least three providers had already been breached by attackers, and user prompts simply do not stay private inside the AI supply chain.
How the traps were set
The "canaries" Moore planted included working access keys tied to empty AWS accounts, passwords for services such as SSH, and unique URLs, DNS names, and private Git repository references. Each secret went into a single prompt sent to a single provider, so any reuse would pinpoint the leak. The bait traveled across the AI services network, in particular through "AI routers" such as OpenRouter and Vercel: they route user prompts to inference engines by price or task fit, and while prompts are encrypted in transit, they are decrypted at every hop of the resale network, because a router must read the plaintext to pick an upstream provider.
Why privacy flags could not stop it
At one breached provider, attackers stole every request sent during a few short windows in late October; a few hundred AWS keys were reused repeatedly and called tens of thousands of times through the AWS API. The attackers browsed S3 buckets and secrets managers, tried to enable hosted models in Bedrock, and added IAM users to keep access. Some stolen keys came from requests the provider had rejected before any model ran, showing the attacker was copying requests at the provider's "front door". In the other two breaches, some stolen requests were marked Zero Data Retention (ZDR). An unnamed Chinese AI lab scraped more than half of the secret URLs sent to its models between July and October, about two-thirds of them marked "no data collection" or ZDR; another US hosting provider was still crawling the same secret URL more than two months after a single call.
The people hit hardest are developers and companies that put API keys, internal documents, and customer data into prompts: once a prompt leaves your own machine, it may pass through routers, resellers, and hosting providers, and it can be read in plaintext at every hop. Privacy flags such as ZDR cannot stop an attacker who is already inside a provider, and cheap resellers may lure traffic with low prices, turn on request logging, and wait to harvest. The report does not name the three breached providers or the Chinese lab, and carries no on-the-record response from the companies; on the same day, OpenRouter emailed customers to warn them that some prompts sent to MiniMax may have been read, and it removed the "Zero Data Retention" label from MiniMax model endpoints.
Moore's advice is blunt: the safest option is to self-host models on premises, though that costs more; the next best is to contract a first-tier provider directly and choose the upstream yourself, rather than accepting a default provider or reseller pool. Either way, stop putting real keys and internal links into prompts.