AI news ·
AI providers leak millions of sensitive secrets through compromised routers and privacy failures
A security researcher planted 2 million digital tripwires across AI services and found at least three hosting providers actively breached, with attackers stealing and reusing hundreds of AWS keys within hours.

A security researcher planted 2 million digital tripwires across a sprawling network of AI services and watched as highly sensitive credentials leaked through the ecosystem "like a sieve." The experiment, designed by Dragos chief research officer HD Moore, revealed that at least three AI hosting providers were actively compromised by attackers siphoning prompts, multiple resellers logged and passed along private data, and several providers ignored end-user privacy flags such as Zero Data Retention.
Moore assembled a collection of functioning Amazon Web Services access keys tied to empty accounts, SSH passwords, unique URLs, DNS names, and private git repository references. Each secret was seeded into exactly one prompt sent to a single provider. If any credential was later used, the canary tactic allowed Moore to trace precisely which link in the chain was responsible. "We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced," Moore wrote. "What we found revealed security breaches, prompt theft, and widespread failures to protect customer data."
How the inference supply chain leaks data
Modern AI applications rarely connect directly to a model from OpenAI or Anthropic. Instead, developers use AI routers such as OpenRouter or Vercel that receive prompts and route them to an available inference engine based on price, latency, or task suitability. While prompts are encrypted in transit, they are decrypted at each hop inside this reseller network. The prompt is read in plaintext by the router, then often passed to a pool of models running inside different hosting providers. Each player in this chain is likely to log the prompt and store it according to its own retention policies.
The most concerning finding involved three separate hosting providers that were breached. Attackers who compromised one provider reused "a couple hundred" of the AWS keys. Moore wrote:
"Each key was sent in exactly one request to one provider, all within a few short windows in late October, and every request we sent during those windows was stolen. Some of the stolen keys came from requests that the provider rejected before a model ever ran-this tells us that the attacker was copying requests at the provider's front door, and not somewhere deeper in the inference stack. In the other two breaches, some of the stolen requests were marked as ZDR. A privacy flag doesn't help when the attacker is already inside."
The parties behind these breaches tested each key, sometimes within hours, probing S3 buckets, secrets managers, and parameter stores for additional credentials. They attempted to enable hosted models in AWS Bedrock and added their own IAM users to maintain access. The stolen bait keys generated "tens of thousands" of AWS API calls. Secrets sent to two of the hacked providers were used together almost immediately from the same rented server, indicating the same adversary likely compromised both.
Resellers logging prompts and scraping URLs
Some leaks originated not from external attackers but from the resellers themselves. To attract prompts, certain resellers offered low prices, then enabled request logging and waited. Moore identified them by sending requests to the same model through different sellers. When credentials were reused after passing through one seller but not another, the reseller was the source.
The experiment also showed that some providers scrape secret URLs embedded in prompts, violating their own privacy terms. Moore detailed one case: "The clearest example involves a Chinese AI lab. This lab's systems fetched the secret URLs from more than half of the prompts that we sent to its models between July and October, including roughly two-thirds of the requests marked as 'no data collection' or Zero Data Retention. The scraper typically showed up two or three days after the request."
Prompts sent to US hosting providers that resell this lab's models triggered the same scraper, meaning the data flowed back to the lab regardless of where the customer thought their data would stay. Later, a git client on a different cloud began attempting to clone the private repositories referenced in prompts that used this lab's models. One US hosting provider crawled a secret URL more than two months after a single inference call. Another ran a scraper that returned for URLs days after each request, no matter which AI router was used.
Minimizing exposure in a leak-prone supply chain
Moore said the most effective defense is hosting models on-premises, though that option carries significant cost. For organizations that cannot self-host, the next best measure is receiving service directly from a first-tier provider that offers "one contract and a privacy policy to read." Even then, there are no guarantees, but the approach reduces the number of parties handling plaintext prompts. AI users should also avoid accepting default provider or reseller pools and instead select upstream providers themselves.
For professionals managing sensitive data across healthcare, legal, or IT environments, the research underscores a structural problem: every intermediary in the inference pipeline represents a potential leak point. Understanding that chain-and actively minimizing it-is now a practical requirement for any team integrating AI into workflows that touch protected information. Organizations investing in AI Security Analytics Courses can build the skills to audit these supply chains and enforce stricter data handling policies internally.
Why this matters for IT, legal, and healthcare professionals
The findings mean that secrets placed into AI prompts-API keys, repository references, internal URLs-face exposure not just from the model provider but from every router, reseller, and intermediary in between. For legal and healthcare teams subject to data residency or confidentiality obligations, the practice of routing prompts through the cheapest available inference engine creates compliance risk that standard vendor assessments rarely capture. IT leaders who rely on third-party AI integrations should verify whether their prompts pass through reseller networks and, as Moore advises, insist on direct contracts with first-tier providers who can document exactly where plaintext data travels. For CIOs building AI strategy, AI IT Strategy Training offers frameworks for evaluating these supply chain risks before they result in a breach.