Proprietary AI Tools

There Are More Hands on Your AI API Call Than You Think: OpenRouter Drops Three Inference Providers and Strips Zero-Data-Retention Status From 125 Endpoints

November 23, 20260 views阅读中文原文
There Are More Hands on Your AI API Call Than You Think: OpenRouter Drops Three Inference Providers and Strips Zero-Data-Retention Status From 125 Endpoints

The short version: On November 22, AI model routing platform OpenRouter published a notice confirming it had removed three inference providers from its platform and revoked the Zero Data Retention (ZDR) eligibility of 125 endpoints. Some of those removals came after an investigation confirmed problems; others were precautionary, taken before anything was proven.

That matters because for the past two years the entire conversation has been about which model is cheaper. This notice is the first time a platform has publicly conceded the other half: the prompt you send passes through more hands than the one you picked, and not every one of those hands intends to honor the privacy switch you set.

First, what an AI router actually is

The "one API, hundreds of models" experience most people use runs on a routing layer. It works a lot like a delivery app:

  • You place an order (send a request), and the platform dispatches it to a shop that can fill it (a provider).
  • Some platforms are themselves aggregators, with other aggregators behind them, layer on layer.
  • Some model developers serve requests directly but also hand traffic to third-party data centers to run.

The problem is that every layer in that chain can read the full prompt and the full response. OpenRouter's notice does not dodge this. It states clearly that the findings "did not show any breach or unauthorized access of OpenRouter" — while acknowledging that the trouble starts after it forwards the request onward.

Think of asking a friend to buy something abroad. The friend asks another friend, who asks another. The item arrives, but your shopping list has been photocopied a few times and nobody told you.

What OpenRouter actually did

The notice lists five actions. The first two are the hard ones and deserve to be read verbatim:

  1. Off-boarded three providers from OpenRouter, based on the results of its investigations.
  2. Removed 125 endpoints from Zero Data Retention eligibility. That includes endpoints flagged by the researchers plus additional endpoints removed "as a precautionary measure."
  3. Adapted the researchers' testing approach so OpenRouter can run those checks on an ongoing basis.
  4. Began technical verification of network paths to identify where requests and responses may be forwarded to third parties.
  5. Will require "appropriate assurances and validation" before restoring ZDR eligibility to any endpoint removed in the review.

The phrase "precautionary measure" in item two deserves its own paragraph. It means part of that 125 was not proven bad — the platform judged it suspicious enough to pull first. Which cuts both ways: the true number of affected endpoints could be higher than 125, or lower. The notice does not break the figure down, and that boundary should be preserved rather than smoothed over.

The research behind it, and why the details sting

OpenRouter's notice credits security firm Dragos, and that report is where this all starts. The method is simple and slightly brutal.

The researchers built a large set of "canary tokens" — unique URLs, hostnames, and AWS keys that had never existed anywhere else — and mixed them into otherwise ordinary prompts sent to AI services. A token can only "fire" one way: somebody actually visits it. If it fires, someone saw the original prompt.

The run spanned July through November 2026 and roughly two million prompts. The findings stack up in three layers, each worse than the last.

Layer one: active theft, in real time. At least three providers had data exposed, and the researchers watched their seeded test prompts get used against controlled test systems and canary AWS accounts. The largest incident involved a couple hundred AWS keys. One detail is the most telling: some of the stolen keys came from requests the provider had rejected before a model ever ran. That places the attacker at the provider's front door, copying traffic on the way in, not somewhere deep inside the inference stack. The door check was effectively decorative.

Once a key was in hand, the playbook was mechanical: test whether it was live, in some cases within hours; list S3 buckets and search Secrets Manager and Parameter Store; then try to enable hosted models in AWS Bedrock, create Bedrock API keys, and add their own IAM users to keep access. The bait keys were used for tens of thousands of AWS API calls. One key seeded on October 21 was still in use three weeks later.

Layer two: the middlemen did not have to break anything. Some marketplace sellers reselling inference capacity only had to list a model at the lowest price, turn on request logging, and wait. Traffic arrived on its own, because the marketplace routes to whoever is cheapest. The researchers watched individuals on what looked like home broadband and personal computers work through their planted credentials one login attempt at a time — including one person on a Mac in France typing root SSH password guesses by hand.

The line worth remembering from the report: a privacy flag only binds the first hop. Your ZDR switch is pressed at the platform's front door. Two or three hops later, nobody presses it again on your behalf.

Layer three: ZDR was widely ignored. Two contrasts stand out.

  • A Chinese AI lab's systems fetched the secret URLs embedded in more than half of the prompts sent to its models between July and October, including roughly two-thirds of the requests marked "no data collection" or ZDR. Dragos did not name the lab. And this was not limited to direct requests: several U.S. hosting providers resell that lab's models, and prompts sent to those U.S. providers triggered the exact same scraper. Customers who deliberately chose a U.S. provider to keep data in the U.S. did not get what they paid for.
  • In another case, a link from a prompt sent to a Chinese provider surfaced in a Feishu (known outside China as Lark) internal chat 24 days later. Within two minutes it was opened from three desktops and three phones in Beijing and Shanghai. Dragos's disclosure to that provider included only the first eight characters of the link — so whoever shared it was reading from a stored copy of the full prompt, more than three weeks after the request.

The big U.S. providers were not clean either. One was still crawling a secret URL more than two months after the researchers used it in a single call to its search endpoint.

What the providers themselves admit

This is the hardest part of the story: it is not only the researchers talking. Providers named in the report published their own notices.

Novita, on November 22, confirmed that a routing configuration issue caused an API forwarding path intended for specific customers to apply more broadly. As a result, MiniMax requests sent through Novita between May 28 and November 11, 2026 may have had their contents accessed or retained by a third party — including requests made directly through Novita and requests routed to Novita via OpenRouter. Novita says it does not itself retain customer prompts or outputs, and that it is following up with MiniMax on retained data and deletion. It also gave a blunt piece of advice: if your requests included live credentials, API keys, or other secrets, treat them as compromised and rotate them.

Phala published its preliminary report earlier, on November 10, but the covered window overlaps this incident: requests for DeepSeek V4 Flash routed to Phala between October 21 and November 8 should be treated as exposed. Phala's account is more severe in one respect — the attacker obtained application credentials and may also have exposed secrets derived from its key management system, affecting some confidential virtual machine tenants. Phala says it contained the incident on November 8, rotated its KMS root key on November 6, and will publish a full root cause analysis.

Line those windows up against Dragos's July-to-November research period and they interlock almost exactly.

Who it is for: what this means in practice

If you occasionally call an aggregator API for small tools: the exposure is limited, but there is one thing to do now — check whether your prompts ever carried real keys, internal hostnames, or database connection strings pasted in out of convenience. Anything that goes into a prompt is handed to every hop.

If you build agents or automated pipelines: you are the exposed group. An agent's defining trait is that it sends whatever it can read, which means environment variables, repo credentials, and config files get packaged into context. Dragos's advice lands squarely here: plant a fake credential in the context your agents already send, and treat any use of it as an incident. A free Canarytoken is enough to start.

If you procure AI services under a contract that says data must not leave the country: this report offers a very concrete lesson — the provider's registered location is not where your data actually stays. Resale, subcontracted inference, and undisclosed hops can route it around the world and back. Either pick your upstream providers yourself instead of accepting the default pool, or go direct with one company, one contract, one privacy policy.

How to use this: four steps to take now

  1. Go through the last three months of prompts and logs and look for real credentials that got pasted in. Rotate anything you find; do not wait.
  2. Plant canary tokens in your agent context. It is the cheapest, fastest-working detection available, and a free Thinkst Canarytoken is enough to begin.
  3. Do not run sensitive work through lowest-price marketplaces. The report's underlying point is simple: if a price is implausibly low, someone else in the chain is making a decision for you.
  4. Start with the first step: treat ZDR as a first-hop promise, not an end-to-end guarantee. Turning it on is necessary. It is not proof that nobody read your data.

Pitfalls, and when not to overreact

Three things are worth keeping straight.

First, there is no evidence here that OpenRouter's own systems were breached. The notice says so explicitly. What was breached sits downstream of its forwarding.

Second, not every provider is compromised. Three were off-boarded and 125 endpoints lost ZDR eligibility, while the platform's total endpoint count is far larger. The real problem is that from the outside you cannot tell which is which — which is exactly why choosing your own upstream and testing for yourself matters.

Third, phrases like "a Chinese AI lab" and "one Chinese provider" are deliberately anonymous in the Dragos report. Any claim mapping them to a specific company name is not something that report says. This article does not make that inference either.

Background

OpenRouter is a commercial, closed-source routing platform that makes money from subscriptions and usage-based billing. It does not train models itself. The value proposition — one key that reaches hundreds of models — is also the cost: a one-to-one trust relationship becomes a one-to-many relationship with an opaque chain in the middle. This notice is worth a standalone article precisely because it drags that normally invisible middle layer into the open.

To be explicit about sourcing: the enforcement actions, time windows, and provider admissions come from the official notices published by OpenRouter, Novita, and Phala. The research process and findings come from the Dragos report. The Decoder's coverage was used to cross-check publication timing and key facts. This article is compiled from those primary documents rather than from a single media retelling.

Sources

  • OpenRouter, "Data Processing Incidents on Inference Providers": https://openrouter.ai/blog/announcements/data-processing-incidents-on-inference-providers/
  • Dragos, "AI Supply Chain Security: What Canary Tokens Found": https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens
  • Novita, "Privacy Notice Regarding MiniMax Requests on Novita": https://novita.ai/blog/minimax-requests-privacy-notice/
  • Phala, "Phala November Preliminary Incident Report": https://phala.com/posts/phala-november-preliminary-incident-report
  • The Decoder, "Security researchers show how cheap AI inference providers steal keys and snoop on users": https://the-decoder.com/security-researchers-show-how-cheap-ai-inference-providers-steal-keys-and-snoop-on-users/

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...