2-Million-Prompt Canary Experiment Reveals Breaches at 3 AI Providers and Stolen AWS Keys
Say you put a working AWS key in a prompt, and an AI router like OpenRouter passes it down the chain. HD Moore of Dragos sent nearly 2 million prompts containing canary keys and watched where they would surface. At least three providers were breached, and some resellers stole secrets from their logs. The Zero Data Retention (ZDR, "do not store data") flag did not help. Don't put real keys in prompts.
AI-processed from Ars Technica; edited by Hamidun News
Picture an ordinary situation: a developer builds an AI app, and a working AWS key or an SSH password accidentally ends up in one of the prompts. To avoid locking into a single model vendor, the app sends requests through an AI router such as OpenRouter or Vercel, which decides where it is cheaper or which model fits best. Researcher HD Moore, principal researcher at Dragos, checked what happens to such secrets next.
He sent almost 2 million prompts containing single-use "traps" (in security they are called canaries, after the birds once used in mines). These are fake but working secrets, and any use of them tells him immediately who stole them. The result is alarming: at least three providers were breached, some resellers stole prompts, and privacy flags like Zero Data Retention (ZDR, "do not store data") did not protect anything.
If you send anything sensitive to an LLM, this affects you directly.
How the trap was built
Moore assembled AWS access keys to empty accounts, SSH passwords, unique links, DNS names and addresses of private git repositories. None of the secrets had been used anywhere before, and the server that received any requests made with them was controlled by the researcher himself.
He inserted each secret into exactly one prompt sent to one provider. If a secret surfaced anywhere, it was clear which link in the chain had leaked it.
Encryption in transit does not help here. A router has to read the prompt in plain text to pick a suitable model. From there, the request may go to a pool of models on different hosts or to another router, for example from Vercel to OpenRouter. At every step it may be written to a log and kept for as long as company policy allows, or even indefinitely.
What he found
The most serious finding was the breach of three separate hosting providers. At one of them, attackers used a "couple hundred" AWS keys. Each key went out in exactly one request, all submissions fell into short windows in late October, and every request in those windows was stolen. Some of the keys were taken from requests that the provider had rejected before the model even ran. That means the attacker was copying traffic right at the "entry point," not deep inside the infrastructure.
In the other two cases, the stolen requests included ones flagged ZDR. A privacy flag is useless if the attacker is already inside.
Parties tied to the breaches tested the keys, sometimes for several hours after they were sent. They browsed S3 buckets, secrets managers and parameter stores looking for new credentials, tried to enable models in AWS Bedrock, and added their own IAM users to keep access. The bait was used "tens of thousands" of times in AWS API calls. Secrets sent to two of the breached providers were used almost immediately from a single rented server, so one adversary was probably behind both breaches.
A separate group is the resellers. They attracted customers with low prices, turned on request logging and waited. Moore identified them by sending requests to the same model through different sellers. If a key was later used only after being sent through one particular seller, that seller was to blame, not the model's developer.
Finally, some providers violated their own privacy terms. This is clearest in the case of a Chinese AI lab: its systems opened secret links from more than half of the prompts sent between July and October. Among requests flagged "no data collection" or ZDR, this happened in about two thirds of cases.
The scraper usually showed up two to three days later. It also fired at US hosts that resell these models, meaning the prompts flowed back to the lab. A customer who chose a US provider to keep data in the US did not get what they paid for.
Later, a git client in a different cloud tried to clone the private repositories from the prompts, but only those sent to this lab's models. One US host accessed a secret link more than two months after the single call. Another ran a scraper that came back for links a few days after every request, regardless of the router.
What to do in practice
The most reliable protection, according to Moore, is to run models on your own servers. That is expensive. The next option is to work directly with a top-tier provider that has "one contract and one privacy policy you can read." There are no guarantees there either, but the chance of a leak is lower. Don't accept provider pools and resellers by default: choose the end providers yourself.
For teams the rule is simple: never send real keys, passwords or tokens in a prompt. Use placeholders and fetch the real values on your own side. If a key does end up in a prompt, treat it as compromised and reissue it. The same logic applies to trade secrets: the ZDR flag remains a promise, not a technical barrier.
- Nov 20, 2026Meta reconsiders its 2030 net-zero target after emissions jump 40%
- Nov 20, 2026Amazon Textract Insights parses bundles of up to 3000 pages and shows the source of every field
- Nov 20, 2026Checking AI-Written Code With Math Gets Cheaper: Opus 5.5 Uses 40% Fewer Tokens
- Nov 20, 2026Anthropic's IPO Valuation Tops $2 Trillion — 3.5 Times the Entire Russian Stock Market
Key Questions Answered
How the trap was built?
Moore assembled AWS access keys to empty accounts, SSH passwords, unique links, DNS names and addresses of private git repositories. None of the secrets had been used anywhere before, and the server that received any requests made with them was controlled by the researcher himself. He inserted each secret into exactly one prompt sent to one provider. If a secret surfaced anywhere, it was clear which link in the chain had leaked it. Encryption in transit does not help here.
What he found?
The most serious finding was the breach of three separate hosting providers. At one of them, attackers used a "couple hundred" AWS keys. Each key went out in exactly one request, all submissions fell into short windows in late October, and every request in those windows was stolen. Some of the keys were taken from requests that the provider had rejected before the model even ran. That means the attacker was copying traffic right at the "entry point," not deep inside the infrastructure.
What to do in practice?
The most reliable protection, according to Moore, is to run models on your own servers. That is expensive. The next option is to work directly with a top-tier provider that has "one contract and one privacy policy you can read." There are no guarantees there either, but the chance of a leak is lower. Don't accept provider pools and resellers by default: choose the end providers yourself. For teams the rule is simple: never send real keys, passwords or tokens in a prompt.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.