Vibeleaderboard
← All Intel
Intel / article

Data Processing Incidents on Inference Providers

Source
openrouter.ai
Author
sdko
Date

Context

OpenRouter says none of its own systems were breached, but some of the hosting companies it sends requests to leaked or mishandled customer prompts, and it has acted on that. OpenRouter lets developers reach many AI models through one API, and for each request it chooses which provider serves the model. In its incident post, OpenRouter says it has removed three providers, withdrawn Zero Data Retention eligibility from 125 endpoints and notified customers who may have been affected. Some of those endpoints were flagged by researchers and the rest were removed as a precaution. The post does not say which three providers were removed.

The findings came from the security firm Dragos. Its research team, led by HD Moore, sent about two million prompts to AI services between July and November 2026, each seeded with canary tokens: URLs and real AWS keys tied to empty, fully logged accounts that appear nowhere else, so any later use shows that someone read the prompt. Dragos reports evidence that attackers took data from at least three providers, and that in two of those breaches some of the stolen requests had been marked for zero retention. It also recorded companies fetching URLs from prompts whose privacy settings should have kept those companies from seeing them. Dragos's post does not name the breached providers.

Three providers have published notices, and they describe different failures. Novita says a forwarding meant for specific customers applied more broadly, so some requests to MiniMax models were sent to MiniMax's own API instead of being processed on Novita's servers. It warns that requests made between May 28 and November 11, including ones routed through OpenRouter, may have been accessed or retained by a third party. Phala says an attacker took control of an inference machine on a new GPU server that lacked its full security hardening, exposing requests to DeepSeek V4 Flash 0731 that OpenRouter routed there, and tells users to treat contents sent between October 21 and November 8 as exposed. Together AI says an AI committed a to one of its public GitHub repositories, and someone used it to download prompt samples from June and July that the company had stored for software testing, then searched them for credentials; five of those were Dragos canaries. OpenRouter says other affected providers have not posted publicly and that it notified affected users regardless.

On OpenRouter, Zero Data Retention limits routing to endpoints whose policy is not to store prompts and responses. That is a commitment about the provider's own storage, and it cannot stop a request being forwarded to another company or read on a compromised server, which is what the Novita and Phala notices describe. Together's leak was different: it involved prompt samples the company had stored for testing, and Together has since made zero data retention the default for new accounts. Dragos makes the broader point that a privacy flag does not help once an attacker is inside. OpenRouter says it is adapting Dragos's method to run these checks continuously, verifying network paths to find requests passed on to third parties, and will require assurances and validation before restoring eligibility to any removed endpoint. Phala and Novita both advise anyone whose requests fell in the affected windows to rotate keys, passwords or other secrets those requests contained.

Terms in this piece · Glossary
  • Secret — A value such as a password, key or token that lets its holder act as someone; it must be kept out of code and shared files.
  • Path — A description of a file or folder location, written using directory names and separators.
  • Agent mode — A mode in which one agent can inspect, edit, run commands, and iterate toward an outcome.
  • Credential — Something presented with a request to prove that the sender is a particular identity, such as a password or a generated key.
Read the source openrouter.ai
Recommended reads
Comments

Checking sign-in…

Loading comments…