Advertisement
Advertisement
Advertisement
22 November 2026·8 min read·By Beatrice Novak

AI supply chain leak exposes secret prompt data

A study on the AI supply chain leak showed prompts were siphoned by attackers, logged by resellers, and scraped by labs.

AI supply chain leak exposes secret prompt data

AI supply chain leaks mount. Security researcher HD Moore, who's chief research officer at security firm Dragos, planted nearly two million trackable secrets into requests sent through a broad network of AI tools. Privacy guarantees don't hold. And millions of digital tripwires deployed across artificial intelligence services reveal widespread data exposure and security compromises, uncovering an ecosystem that exposes sensitive user prompts to unauthorized third parties, resellers, and active attackers.

Modern application architectures don't usually connect a user directly to a single model provider, so developers rely heavily on AI routers or AI gateways instead. Services such as Vercel and OpenRouter act as brokers, accepting prompts from end users and forwarding them to available inference engines based on factors like price or technical fit. This avoids giant provider lock-in. But it's introducing multiple middle layers into the data path.

Decryption at every hop creates massive risk

Requests move across the internet using standard transport encryption, but that protection drops the moment a request reaches an intermediary. The prompt must be decrypted. It's turned into plain text at every stop along the route. An AI router needs to read the unencrypted text to choose the right inference provider, and the payload may then pass through additional routers or pools of hosting providers running the actual models. So each party in this relay can log and store the text indefinitely under its own data retention rules.

To measure how far this unencrypted data travels, Moore created a massive collection of unique, functioning access credentials tied to empty Amazon Web Services accounts, adding passwords for services like SSH. He also included unique URLs, DNS names, and references to private git repositories. They'd never been used before. And it's Moore who personally controlled the infrastructure designated to receive any incoming connections.

He seeded exactly one secret. And if any credential showed activity on his servers from that single prompt sent to a specific provider, it's the canary that immediately exposed the exact link in the chain that leaked it.

Attackers siphon data straight from provider front doors

They're already being exploited. The findings showed that systemic vulnerabilities exist in the wild, where an experiment revealed three distinct hosting providers that had been completely compromised by external attackers. So, these malicious actors intercepted incoming prompts and extracted the canary secrets embedded inside them.

On one compromised hosting provider, attackers grabbed every request sent during specific late-October testing windows, reusing a couple hundred AWS keys right at the entry point of the service.

Market Context: According to CrowdStrike, valid account abuse was the leading initial access tactic in 35% of cloud incidents in the first half of 2025.
Requests don't reach model engines. So they're copying incoming traffic right at the provider's front door.

Intruders didn't waste any time. After stealing the credentials, they initiated tens of thousands of AWS API calls to probe for extra access while their automated scripts looked through S3 buckets, secrets managers, and parameter stores to steal more keys. They attempted to activate hosted models inside AWS Bedrock. And they tried to create new IAM users to establish permanent access. In two breaches, keys that were stolen from separate providers surfaced almost immediately on the same rented server, proving that a single adversary was behind both intrusions.

It's a second major vulnerability. It stems from low-cost resellers operating in the marketplace. To draw traffic, some providers offered rock-bottom pricing on model access. So, by routing identical requests through different vendors, Moore tracked down the exact resellers who were harvesting input data after turning on full request logging and waiting for credentials to pass through.

Privacy flags fail to stop data collection

Data harvesting didn't stop. And terms like Zero Data Retention (ZDR) or explicit flags requesting no data collection were frequently ignored across the supply chain, even when developers and end users made explicit privacy choices.

  • Chinese AI lab scraping: Systems belonging to a China-based AI lab fetched secret canary URLs from over half the prompts sent to its models between July and October.
  • Bypassing ZDR flags: Roughly two-thirds of the requests scraped by the Chinese lab carried explicit ZDR or no data collection flags.
  • Reseller passthrough: Prompts sent through US hosting providers that resold models from this lab triggered the same scraping, sending user data directly back to the lab.
  • Git repository cloning: A git client running on a separate cloud platform attempted to clone private repositories referenced in prompts sent to the Chinese lab's models.
  • Long-term web crawling: One US hosting provider crawled a secret URL more than two months after it appeared in a single search endpoint request.

It's a broken promise. So, this cross-border flow breaks key geographic promises made to customers when an enterprise picking a domestic provider to keep data inside the US saw its prompt payloads route back to foreign servers anyway.

OpenRouter issues customer warning following investigation

Industry players made quick adjustments. It's all from investigation fallout. And on Nov 21, reports across technical communities revealed that OpenRouter sent notice emails to affected customers, confirming that some user prompts may have been read without authorization.

AI supply chain leak exposes secret

OpenRouter cited the research directly. It's confirmed that potential data exposure involved prompts routed to MiniMax, a model provider based in China, where planted canary URLs were accessed through direct requests or resellers including AtlasCloud, Novita, and StreamLake. So OpenRouter subsequently stripped the Zero Data Retention label from its MiniMax model endpoints.

Security failures in third-party inference layers leave enterprise buyers with very few reliable options for protecting proprietary text.

We've sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced, Moore wrote when detailing what they found. Prompt theft occurred. And what they found revealed security breaches alongside widespread failures because providers can't protect customer data.

How developers can lower their exposure

Fixing these leaks isn't quick. So standard transport layer security doesn't prevent intermediary nodes from viewing, logging, or losing control of payload contents across this multi-tiered architecture, while relying on default marketplace routing pools leaves organizations vulnerable to rogue sellers and compromised nodes.

Self-hosting offers maximum defense

Hosting models on dedicated on-premises infrastructure provides the strongest protection against plain-text interception because running them in-house removes external middlemen, eliminates intermediary logging, and keeps prompt payloads entirely inside the local network boundary. But hardware requirements and compute overhead make self-hosting financially unfeasible. Many smaller organizations can't.

Direct provider contracts reduce chain length

For teams that can't run local hardware, contracting directly with a first-tier model vendor is the next best alternative. Dealing directly with a primary provider relies on one clear contract and a single privacy policy, which avoids complex broker layers while reducing the total number of hops where plain text can be siphoned. Zero data leakage isn't guaranteed. So developers should also manually select their upstream providers rather than accepting automatic reseller pools.

Frequently Asked Questions

How did researcher HD Moore uncover the AI supply chain leak?

HD Moore planted nearly two million trackable secrets—such as AWS credentials, passwords, and private URLs—into requests sent through a broad network of AI tools. He monitored infrastructure he personally controlled to track any incoming connections triggered by those canary secrets. When a planted credential showed activity, it immediately exposed the exact link in the chain that leaked the prompt data.

Why does the architecture of AI routers and gateways create security risks for prompt data?

Developers rely heavily on AI routers and gateways to avoid lock-in, which introduces multiple middle layers into the data path. Standard transport encryption drops the moment a request reaches an intermediary because the prompt must be decrypted into plain text for the router to select an inference provider. Consequently, every party in the relay can log and store the unencrypted text indefinitely according to its own data retention rules.

What actions did external attackers take after intercepting credentials from compromised hosting providers?

Attackers completely compromised three hosting providers and siphoned incoming prompts right at the service entry points to steal embedded AWS keys. Upon stealing the credentials, the intruders executed tens of thousands of AWS API calls to search S3 buckets, secrets managers, and parameter stores for extra access. They also tried to activate hosted models in AWS Bedrock and create new IAM users to establish permanent access.

What did OpenRouter disclose and change following the investigation into data exposure?

On November 21, OpenRouter issued notification emails to affected customers confirming that user prompts routed to China-based provider MiniMax may have been read without authorization. The exposure occurred through direct requests as well as model resellers like AtlasCloud, Novita, and StreamLake. Following the investigation fallout, OpenRouter stripped the Zero Data Retention label from its MiniMax model endpoints.

How can organizations reduce their exposure to prompt data harvesting and leaks?

Developers can achieve maximum defense by self-hosting models on dedicated on-premises infrastructure, which removes third-party middlemen and keeps prompt payloads within the local network boundary. Organizations unable to afford self-hosting should contract directly with primary model providers to avoid complex broker layers and reduce total relay hops. Developers should also manually select upstream providers rather than accepting automatic reseller pools to limit exposure to rogue vendors.

Beatrice Novak
Written by
Business and Technology Editor

Beatrice Novak covers the business of technology, from enterprise software and cloud platforms to the strategy behind the biggest deals. She follows how companies adopt new tools and what it means for the wider economy.

💬 Comments (0)

Sign in to leave a comment.

No comments yet. Be the first!

Advertisement