Two million booby-trapped prompts: Dragos uncovers leaks at AI providers, OpenRouter drops three hosts
Security firm Dragos planted unique secrets in about two million prompts sent to AI services between July and November. The result: at least three hacked hosting providers, resellers reading requests, and "zero data retention" settings ignored. OpenRouter dropped three providers, and Phala, Together and Novita published their own incidents. Here are the affected windows and the precautions that matter.

What happens to a prompt once it is sent to an AI service? To find out, HD Moore, chief research officer at security firm Dragos, sent about two million booby-trapped prompts between July and November 2026. Each contained “canaries”: a URL, a domain name and an AWS key that were unique and never used anywhere else. If someone uses one, they have read the prompt. His study published on November 19 describes three kinds of leaks, and several providers have since acknowledged incidents.
Three findings, from hacking to ignored settings
1. Hacked hosting providers. Dragos found attackers stealing requests from at least three providers, none of which knew. The first incident dates back to mid-July. In the most serious case, about two hundred trap AWS keys were sent in late October: every request in those windows was stolen, including requests rejected before the model ran. The attacker was therefore copying traffic at the front door. The keys were sometimes tested within hours, then used to search for data and try to set up lasting access: tens of thousands of AWS API calls in total. On November 1, the attacker added a penetration-testing framework driven by a language model to its tools.
2. Resellers reading requests. On an inference marketplace, sellers offer access to a model at the lowest price. All they need to do is turn on logging to read the prompts. Dragos saw individuals, on home connections, try to log in to its trap servers, including one in France typing by hand the passwords found in two prompts.
3. Privacy settings ignored. The systems of an unnamed Chinese AI lab visited the secret links in more than half of the prompts sent to its models, including about two thirds of the requests marked Zero Data Retention (ZDR, the setting meant to forbid any storage). U.S. hosting providers reselling its models triggered the same crawler. In another case, a secret link was posted in a company chat in China 24 days after the request, and opened from Beijing and Shanghai.
What the providers acknowledge
| Who | What happened, according to the provider | Requests affected |
|---|---|---|
| OpenRouter (router) | 3 providers removed, 125 endpoints stripped of the ZDR label, customers notified; no breach at OpenRouter | per the notices sent to customers |
| Phala | takeover of a poorly protected inference server | DeepSeek V4 Flash 0731 via OpenRouter, from October 21 (00:00 UTC) to November 8 (05:35 UTC) |
| Together AI | credential committed to GitHub by an AI coding agent, prompt samples downloaded | MiniMax M3, GLM-5.2, Kimi K2.6, mainly June 24 and July 4, 8 and 14 |
| Novita | requests mistakenly forwarded to MiniMax’s official API | MiniMax models, from May 28 to November 11 |
Together AI’s case joins an already long list: after the Railway breach and the Cursor incident, a poorly protected credential once again opened access to customer data, this time committed by a coding agent. Together says it has made ZDR the default setting for new accounts.
Dragos names neither the hacked providers nor the Chinese lab, and OpenRouter does not name the three providers it removed. Nothing allows us to say they are the same.
What you can do
- If you used the models and periods in the table, change every key, password and token present in your prompts: that is Phala’s and Novita’s advice.
- Keep secrets out of agent context. A coding agent sends everything it reads,
.envfiles included. - Choose your providers. With a router, set the list of hosts yourself rather than accepting the cheapest. Dragos points out that a privacy setting “only binds the first hop” of the chain.
- Plant your own canaries: a fake key in your agents’ context, any use of which signals a leak. Dragos mentions Thinkst’s free Canarytokens.
- For the most sensitive work, host it yourself. Dragos gives benchmarks: an NVIDIA DGX Spark at $6,950, a 512 GB Mac Studio from $22,299.
Our take
Dragos’s study changes the nature of the debate on model privacy. The risk no longer lies only in a vendor’s policy, but in the whole chain: routers, resellers, hosts, and sometimes the model maker itself. The ZDR label, presented as a guarantee, has turned out to be declarative at best. OpenRouter’s response, removing 125 endpoints from the label and promising to run these tests continuously, shows that the canary method can become a routine control. For a company, the consequence is immediate: treat any prompt sent to a third-party service as potentially readable by others, and strip secrets from it before sending. The names of the hacked providers are still unknown, and that is the next piece of information to watch.
Sources
- Dragos — Three breaches, endless leaks: Your AI supply chain needs canaries (HD Moore, November 19, 2026) Primary source
- OpenRouter — Data Processing Incidents on Inference Providers (November 22, 2026) Primary source
- Phala — November Preliminary Incident Report (November 10, 2026) Primary source
- Together AI — Security event: Prompt leakage for requests in June and July of 2026 (November 21, 2026) Primary source
- Novita — Privacy Notice Regarding MiniMax Requests on Novita (November 22, 2026) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
How we work: editorial charter and methodology · our verification process · corrections log. Spotted an error? Write to contact@neuralbrew.ai.


