AI 4 min read Apex Wire Tech Desk

OpenRouter Cuts 3 AI Providers After Prompt Theft: What Users Must Do

OpenRouter removed three AI providers and stripped 125 endpoints of Zero Data Retention status after researchers caught months of prompt theft at host firms.

Line chart of OpenAI usage on OpenRouter from January 2025 to September 2026, showing a sharp rise in mid-2026.

Image: Wikimedia Commons / FoxTPNL

OpenRouter said on Sunday, November 22, that it has removed three artificial intelligence providers from its platform and stripped 125 endpoints of Zero Data Retention status after researchers showed customer prompts had been stolen or mishandled for months.

The AI gateway said it has notified every customer who may be affected and that the findings showed no breach of OpenRouter itself. The action follows research published on November 19 by Dragos, the cybersecurity firm, which sent about two million prompts carrying unique tracking secrets to AI services.

Key Takeaways

  • OpenRouter off-boarded three providers, removed 125 endpoints from Zero Data Retention eligibility and notified all potentially affected customers.
  • Dragos seeded about two million prompts with canary tokens; attackers stole data at three providers from mid-July.
  • The largest theft involved a couple of hundred AWS keys; sellers logged prompts and a Chinese lab fetched links from over half of its prompts.
  • Novita, Phala and Together AI have published incident notices; users should rotate any secrets sent through affected routes.

Which AI Providers Were Removed by OpenRouter and Why

OpenRouter has not named the three providers it removed, and Dragos has not named the three it found breached. The 125 endpoints include flagged providers plus others cut as a precaution; three providers have published notices, linked from its announcement.

ProviderWhat was disclosedWhat users were told
NovitaA routing fault sent some MiniMax requests to MiniMax's API.Requests went outside the chosen route.
PhalaAn attacker controlled an inference server and could view request data.DeepSeek V4 Flash 0731 requests from October 21 to November 8 are exposed.
Together AIAn unauthorized party downloaded stored prompt samples after a coding agent leaked a credential.Samples came from June and July MiniMax, GLM and Kimi requests.

How Dragos Canary Tokens Exposed Months of Prompt Theft

Dragos planted unique web links and credentials, known as canary tokens, in about two million prompts sent to AI services. A planted secret can only be learned in one place, so any use of it proves someone read the prompt.

The largest breach involved a couple of hundred AWS keys, each sent in exactly one request in late October. Every request in those windows was stolen, including some rejected before a model ran. The keys were tested within hours, then used to list S3 buckets and search Secrets Manager, driving tens of thousands of API calls.

Sellers on an inference marketplace listed models at the lowest prices, enabled logging and collected prompts. In one case, a person in France typed root SSH logins by hand. One seller adds about 1,000 hidden tokens to every prompt.

A Chinese lab fetched secret links from over half of the prompts sent to its models, including two-thirds of those marked Zero Data Retention. One link surfaced in a Feishu chat 24 days later, opened within two minutes in Beijing and Shanghai. It echoes concern about Chinese AI models, after Cursor said it would explore phasing out Chinese open-weight models.

What OpenRouter Users Should Do Now to Protect Prompts and Keys

Anyone who sent prompts through OpenRouter in recent months should rotate API keys, tokens and passwords in those prompts, starting with the routes named in the provider notices. Phala's window runs October 21 to November 8; Together AI's samples came from June and July.

Users should choose their upstream providers and check that privacy settings change where requests go. Self-hosting is the only option in which no third party reads a prompt. As AI agents take on longer tasks, they carry more credentials, and their time horizons have been doubling.

OpenRouter said it will keep updating its list of provider security announcements.

Frequently Asked Questions

Were my OpenRouter prompts exposed?

OpenRouter said it has notified all potentially affected customers. Anyone who sent secrets through OpenRouter routes should still rotate those credentials.

What is Zero Data Retention on OpenRouter?

Zero Data Retention, or ZDR, means prompts and responses are not stored. OpenRouter removed 125 endpoints from ZDR eligibility.

Which three providers did OpenRouter remove?

OpenRouter has not named the three providers it off-boarded, and Dragos has not named the three it found breached, so the two lists cannot be matched publicly.

How do I check if my API keys were exposed?

Check account activity where credentials appeared in prompts, including AWS CloudTrail logs, and rotate any key that passed through an affected route. Seeding canary tokens gives an early warning.

Reported and fact-checked by our editorial desk against independent primary sources. Read our Editorial Policy.
Apex Wire Tech Desk

Artificial intelligence, enterprise tech, gadgets, science, gaming and EVs.

More from the Apex Wire Tech Desk →

Share
✓ Link copied