Cybersecurity

OpenRouter cut three AI hosts after bait prompts exposed months of theft

Adrian Kessler
Add us on Google

OpenRouter, the service that many AI apps use to send prompts to whichever model host is cheapest or fastest, has removed three of those hosts from its platform and stripped its Zero Data Retention label from more than a hundred endpoints. The trigger was an experiment that planted fake credentials in prompts and then watched who tried to use them. Some of the prompts that leaked had been marked as never to be stored.

If you use a chatbot, coding assistant or writing tool built on someone else’s model, your text may pass through several companies before an answer comes back, and each of them can read it in full. OpenRouter says the findings showed no breach of its own systems. The weak points sat further down the chain, at the companies that actually run the models.

The research comes from Dragos, a security firm whose chief research officer, HD Moore, led the work. Over five months his team sent close to two million prompts, each carrying a web link, a server password or an Amazon Web Services key that existed nowhere else. A planted secret can only be learned from the prompt that carried it, so any use of one proves that someone read that prompt. The first tokens fired within days.

The most serious finding was plain theft. Attackers had broken into three inference providers, and none of the three knew until the researchers told them. In the largest case, every request sent to one provider during a few short windows was copied, including requests the provider rejected before a model ever ran, which puts the thief at the front door rather than deep inside. The stolen AWS keys were tested, sometimes within hours, then used to search storage and secret vaults across tens of thousands of API calls. Some of the stolen requests carried the zero-retention flag.

Other leaks needed no break-in. On one reseller marketplace, sellers listed popular models at the lowest price, switched on request logging and kept what flowed through. One person on a home connection in France took passwords from two prompts and tried to log in as root over SSH, typing each attempt by hand. Separately, a Chinese AI lab fetched the secret links from more than half of the prompts sent to its models, including about two thirds of those marked for no data collection, and US hosts that resell its models fed the same scraper. One link turned up in a Feishu work chat, opened from six devices in Beijing and Shanghai, 24 days after a single request.

Three hosts have published notices. Phala said an attacker took control of one of its inference servers and could see requests while they were processed. Together AI said an AI coding agent committed a credential to a public GitHub repository, which let an intruder download stored prompt samples and test the passwords inside them; five of those were the researchers’ bait. Novita said a routing error sent some MiniMax requests on to MiniMax’s own service.

The cleanup has clear limits. OpenRouter has not named the three providers it removed, Dragos has not named the three it found breached, and neither has said whether they are the same companies. One breached provider still had not disclosed its incident publicly, according to the researchers. Dropping a host also cannot recall prompts that were already copied: planted keys and links were still being used weeks after the first thefts. And people using an app built on these gateways usually have no way to see which host handled their text.

Moore’s advice starts from one limit: a privacy flag binds only the first hop, and AI agents will send whatever they can read, credentials included. Self-hosting is the only setup in which no third party reads a prompt. Going straight to a single major provider leaves one contract and one privacy policy to check, and router users should choose the upstream hosts themselves instead of accepting the cheapest pool.

Dragos published its findings on November 19. OpenRouter emailed affected customers and published its incident notice over the following weekend, and Phala tells anyone who sent secrets to one DeepSeek model through the gateway between October 21 and November 8 to rotate them. OpenRouter says it is building the researchers’ bait-prompt tests into its own platform so the checks run continuously, and removed endpoints will regain the zero-retention label only after assurances and validation. Dragos is now working with about a dozen providers.

As Moore put it, where your prompts go is a decision, “and if you don’t make it, the lowest price will make it for you.”

Tags: , , , , ,

Add us on Google

Discussion

There are 0 comments.