Dragos: 3 AI Inference Providers Hit in 2M-Prompt Test Dragos researchers sent about 2 million AI requests containing canary credentials and unique tracking URLs through inference services between July and November 2026 and found at least three AI inference providers had been compromised by attackers who copied incoming prompts, according to TheDecoder's coverage. Some stolen AWS access keys were tested within hours of appearing in a request, with follow-up activity including listing storage buckets, hunting for secrets and attempting to enable Amazon Bedrock models. The test also found some inference-marketplace resellers could read and retain customer prompts through request logging, and that requests flagged with privacy settings such as "Zero Data Retention" or "no data collection" were not consistently handled according to those settings; the affected providers were not named. Dragos researchers sent about 2 million AI requests stuffed with fake credentials and unique tracking URLs through inference services between July and November 2026. At least three AI inference providers turned out to be compromised by attackers who copied the incoming prompts, according to coverage from TheDecoder. The finding turns a quiet assumption on its head: that the company routing your prompt is, at worst, a neutral pipe. The decoy data did not stay decoy for long. Some of the stolen AWS access keys were tested within hours of appearing in a request, and the follow-up activity looked like a practiced playbook: list storage buckets, hunt for secrets, try to switch on Amazon Bedrock models. Readers who followed our earlier report on the Together AI prompt leak https://shattered.io/together-ai-prompt-leak-coding-agent-credential-3-models-4-days/ will recognize the shape of the problem, a credential that rode along inside a prompt and ended up somewhere it never belonged. This analysis walks through what the test showed, what remains unconfirmed, and what it means for teams that route coding agents and chatbots through cheap AI inference providers. Several details are still missing from public reporting, including the names of the affected providers. We flag every gap rather than fill it with guesses. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F What Dragos Tested and What It Found The research comes from Dragos, the industrial cybersecurity firm. Coverage identifies HD Moore as the named researcher, and describes him as Dragos’s Chief Research Officer. The test ran from July to November 2026, which makes it a months-long observation rather than a one-day stunt. The team submitted about 2 million AI requests. Each carried bait: canary credentials that look like real secrets, unique URLs that nobody should ever visit, DNS names, and code-repository links. If any of those items showed up in an attacker’s hands, the unique values told researchers exactly which request leaked and roughly where. The headline result is blunt. At least three AI inference providers were found to have been compromised by attackers who copied incoming prompts. Beyond that, some inference-marketplace resellers could read and retain customer prompts through request logging, including prompts that contained credentials. And requests flagged with privacy settings such as “Zero Data Retention” or “no data collection” were not consistently handled according to those settings, per the reporting. Those are three different failure modes. One is an outside break-in, one is an inside capability, and one is a broken promise. Treating them as a single story about AI providers stealing keys, as some headlines have, blurs what the evidence supports. How the Canary Credential Method Works Canary testing is old and simple. You plant a value that has no legitimate use, then watch for anyone touching it. If a fake AWS key gets tried against AWS, a human or a script copied it from somewhere it should not have been. If a unique URL gets fetched, someone read the text that contained it. Why unique URLs matter A shared bait URL would only prove that somebody saw bait. A unique URL per request lets researchers tie a specific hit to a specific provider and a specific moment. That precision is how a test like this can separate a provider’s own staff or tooling from an attacker who reached in through a compromise. The same logic applies to the AWS keys. A key that exists only inside one prompt, sent to one provider, and then shows up in a login attempt points at a short list of suspects. It also lets a defender measure the delay between exposure and abuse, which is where the finding about tests within hours comes from. What the method cannot prove A canary shows that data left a path. It does not by itself show intent. A provider whose logs were scraped by an intruder and a provider that sold its logs look identical from the bait’s point of view. That limit matters for the next section, because the public sources do not name any provider and do not establish that any of them acted on purpose. Three Compromised Providers, Zero Names The count is the most quotable number in the story: at least three AI inference providers compromised. The identities are not. The available reporting does not name the providers, the marketplace resellers, or any specific router involved, and we are not going to guess. That silence has practical consequences. A developer cannot check a list and cross a name off. Instead, every team that sends secrets through a third-party inference endpoint has to assume the worst case applies to them until they can show otherwise. The standard disclosure practice of notifying vendors first and naming them later may explain the gap, but the sources we reviewed do not say. One more caveat deserves space. Some social posts describe the story as providers stealing keys. The fact sheet behind this report does not support that framing. The documented claim is that providers were compromised and that prompts were copied, or that logged data was exposed. Whether a given operator abused customers on purpose is unconfirmed. AWS Keys: What Attackers Did Within Hours The scale of key theft varies by account. One report says hundreds of AWS keys were harvested, and another puts the total at roughly 200 live AWS keys. Exact counts of stolen keys, AWS API calls, and affected customers are inconsistent across the coverage, so treat any single figure as approximate. The conservative reading is roughly 200 or more. Speed is the part that should worry defenders. Some of the stolen keys were tested within hours of appearing in requests, which means attackers watched the prompt stream in something close to real time. Nobody waits days to sift through a leaked log dump when a live feed is available. The observed activity reads like a reconnaissance checklist. The table below summarizes what the reporting says attackers did with the keys. | Observed action | AWS service | What it suggests | |---|---|---| | List S3 buckets | Amazon S3 | Mapping stored data worth taking | | Search for secrets | AWS Secrets Manager | Hunting more credentials to escalate | | Search parameters | Systems Manager Parameter Store | Looking for configuration and tokens | | Try to enable models | Amazon Bedrock | Seeking free AI compute on someone else’s bill | | Create API keys | Amazon Bedrock | Building a durable foothold for model access | | Add IAM users | AWS IAM | Planting persistence that survives key rotation | Two items stand out. Adding IAM users is a persistence move, because rotating the leaked key does not remove an account the attacker already created. And the Bedrock attempts show a motive that fits the 2026 economy: access to expensive models without paying for them. Teams that want to shrink the blast radius of exactly this pattern can start with our walkthrough on AWS IAM least privilege https://shattered.io/aws-iam-least-privilege-identity-center-access-analyzer-12-steps-2026/ . A key that can only read one bucket is a much smaller prize than a key that can mint new users. A practical wrinkle sits underneath all of this. Many developers never type a credential into a chat window on purpose. The key arrives through a pasted stack trace, an attached configuration file, a terminal log that an agent scooped up, or a code comment someone forgot to remove. Each of those paths is invisible at the moment of sending, and each is the kind of leak a canary test catches well. Cost pressure compounds the habit. Teams running coding agents all day burn through millions of tokens, so a cheaper endpoint feels like a free win. The Dragos numbers suggest the saving can come with a hidden bill, paid later in incident response https://shattered.io/category/security/ hours, rotated keys, and unexpected cloud charges. Resellers and Request Logging Inference marketplaces sell access to many models through one API. Their pitch is price and convenience, and for many teams the pitch works. The Dragos findings add a cost that does not appear on the pricing page: some resellers could read and retain customer prompts through request logging, including prompts that held credentials. Logging itself is normal. Providers log requests to debug failures, bill customers, and fight abuse. The danger starts when logs hold plaintext prompts and the logging layer is weakly protected, widely accessible, or kept longer than customers expect. A log that includes an AWS access key is a credential store nobody designed as one. Resellers also add a hop. A request may pass the marketplace, then an upstream host, then the model operator. Each hop can log. Customers who read only the first company’s privacy page are reading one chapter of a three-chapter contract. When Zero Data Retention Did Not Hold Perhaps the sharpest finding is about promises. Requests marked with “Zero Data Retention” or “no data collection” were not consistently handled according to those settings, based on the reporting. In plain terms, a flag that tells the service not to keep your data did not always stop it from being kept or seen. The word consistently carries weight here. The sources do not say every flagged request leaked, and they do not say how many failed. They say the behavior was uneven. For a privacy control, uneven is nearly as bad as absent, because a customer has no way to know which of their requests fell through. This is also why a canary test beats reading a compliance document. A policy page states intent. A planted credential that gets tried against AWS states what happened. Security teams evaluating vendors should ask whether a provider will allow, or has already passed, independent bait testing of its retention claims. The issue echoes the concerns raised in our piece on prompt injection defense in RAG pipelines https://shattered.io/prompt-injection-rag-defense-scanner-guard-12-steps-2026/ , where untrusted text flows through layers that nobody fully audits. Here the untrusted party sits on the delivery path itself rather than inside the document. The Chinese AI Lab and the Secret Test URLs One detail drew extra attention. A Chinese AI lab accessed secret test URLs in more than half of the requests in the reported test. Those URLs existed only inside the planted prompts, so a visit means someone or something read the prompt text and followed the link. Resist the urge to read geopolitics into that line. The sources do not name the lab, do not say how the lab came to see the prompts, and do not say whether it was a provider, a customer of a provider, or a downstream consumer of logged data. A figure above half is large, but it describes the reported test, not the whole market. The finding does support a narrower point. Prompt text sent to cheap endpoints can end up being processed by parties the user never chose. Whether that is data harvesting for model training, link prefetching, or something else is not established by what has been published. Every Intermediary Sees the Full Prompt The structural lesson comes from reporting that stresses one fact about AI routing: each intermediary handling a request can potentially see the full prompt and the full response. Unlike a web page secured end to end between your browser and a bank, an inference request has to be readable by the service that runs the model. Routers and resellers in the middle inherit that visibility. That design is not a bug in any one product. A model cannot answer a prompt it cannot read. What changes the risk is how many parties sit between the user and the model, and how carefully each one treats what it sees. Agentic coding tools make the exposure larger. An agent that reads a repository, an environment file, and a terminal session can fold secrets into context without the developer noticing. A person who would never email an AWS key to a stranger may let an agent send it across three middlemen automatically. The same trust-chain thinking applies to software dependencies, as the Microsoft Agent Framework PyPI compromise https://shattered.io/microsoft-agent-framework-pypi-compromise-two-releases-33-minutes/ showed in a different corner of the AI stack. In both cases a component developers considered plumbing turned out to be a point of attack. What Is Confirmed and What Is Not Because the story is moving fast and social posts have stretched it, a plain ledger helps. The table separates what the available reporting supports from what it leaves open. | Claim | Status | Detail | |---|---|---| | Dragos ran the test | Reported | HD Moore named, described as Chief Research Officer | | Test window | Reported | July to November 2026 | | About 2 million requests with canary data | Reported | Fake credentials, unique URLs, tracking data | | At least three providers compromised | Reported | Attackers copied incoming prompts | | Roughly 200 or more live AWS keys harvested | Inconsistent | Accounts range from about 200 to hundreds | | Keys tested within hours | Reported | Some, not all, keys | | Provider names | Unconfirmed | Not clearly established in the sources | | Attacker identity | Unconfirmed | Not documented | | Providers intentionally stole keys | Unconfirmed | Compromise or logged-data exposure is the documented claim | Keep that last row in mind when you read summaries. “Cheap AI providers steal your keys makes a sticky headline. The more defensible version says that cheap AI providers, and in some cases their resellers, exposed keys through compromise and logging. Market Impact: Cheap Inference Pays a Trust Discount Inference is a price war. Marketplaces and resellers compete by undercutting first-party API rates, and developers chasing lower token bills have fueled their growth. The Dragos result gives procurement and security teams a concrete reason to ask what that discount hides. We expect the market to respond in three ways, and each is an analyst’s read rather than a reported fact. First, buyers will start asking for proof of log handling, not just a policy page. Second, vendors with strong retention controls will turn them into a marketing line. Third, security-conscious customers will pay more for direct contracts, which narrows the price gap that made resellers attractive. The risk is not confined to small buyers. Large firms already have a pattern of learning about exposure through vendors, a theme we covered in the Uber vendor credential breach https://shattered.io/uber-vendor-credential-breach-4-days-to-confirm-149346-drivers/ report. An inference provider is now one more vendor on that list, with the added twist that it handles the most sensitive thing developers type. Cloud providers feel the effect too. Stolen keys that try to enable Amazon Bedrock models turn AWS customers into involuntary funders of someone else’s AI usage, which is both a billing problem and a trust problem for the platform. Compliance teams will feel this too. Frameworks that ask where regulated data travels were written before prompts became a routine data flow. A vendor questionnaire that never asks about inference routing, prompt logging, and hop counts is now incomplete, and auditors who read the Dragos findings will likely add those lines. Competitive Comparison: Direct API, Reseller, Self-Hosted Teams have three broad ways to run inference, and the Dragos findings shift the trade-offs. The comparison below is qualitative. It draws on the reported risks and on general architecture, not on measured rates for any named product. | Option | Parties who can see the prompt | Main risk from the Dragos findings | Main benefit | |---|---|---|---| | Direct from the model maker | The model maker | One vendor’s logging and breach exposure | Shortest path and a single contract | | Cloud-hosted model service | The cloud provider and sometimes the model maker | Cloud account misconfiguration | Existing enterprise controls | | Marketplace or reseller | Reseller, upstream host, model operator | Extra hops, request logging, uneven privacy flags | Low price and many models behind one API | | Self-hosted open-weight model | Your own infrastructure | Your own patching and access control | Prompts never leave your network | Self-hosting looks attractive after a story like this, and it deserves a fair look. It removes third-party visibility, but it moves the whole security burden onto your team. Our explainer on open-weight versus open-source AI licenses https://shattered.io/open-weight-vs-open-source-ai-4-models-4-licenses-kimi-mistral-qwen-deepseek-2026/ covers the licensing side of that choice, and hardware cost is a separate calculation. Direct access is not automatically safe either. A first-party provider can be breached, and its logs can be mishandled. The Dragos test does not rank any named provider, so it offers no basis to say a direct API is safe and a marketplace is not. It does show that adding hops adds exposure. Historical Context: The Supply-Chain Lesson Repeats Security history is full of trusted middle layers that became the weak link. Package registries, ad networks, managed service providers, and third-party analytics scripts each had their turn. The pattern is consistent: a convenient shared service gains adoption faster than its security review, then a researcher or an attacker finds the gap. AI inference follows the same arc at compressed speed. Two years ago, many teams called a hosted model a novelty. In 2026 they pipe source code, customer tickets, and cloud credentials through it daily. The attack surface grew with usage, and the review process did not keep up. Canary testing has its own long history, from planted documents in intelligence work to honeytokens in modern security operations. What the Dragos test adds is scale and a new target. Two million bait requests across several months is a serious sample, and it aims directly at a layer of the stack that rarely gets audited from the outside. The takeaway from history is not that the technology failed. It is that trust in an intermediary should be earned with evidence, and evidence of this kind was missing until a researcher built it. What Developers and Security Teams Should Do Now The immediate advice splits into prevention and response. Both start from one rule: assume any text sent to a third-party inference endpoint can be read by more than one company. Before a prompt leaves your machine - Strip secrets from prompts. Run a secret scanner on agent context and block requests that contain key patterns. - Prefer short-lived credentials. A token that expires in an hour has little value to someone who finds it a day later. - Scope keys tightly. Give each key the smallest set of permissions its job needs. - Ask vendors in writing how they log, how long they keep logs, and who can read them. Request evidence, not just a policy link. - Seed your own canaries. Plant a fake key and a unique URL in a test prompt and watch for any use. If you think a key already leaked - Rotate the key at once, then check for new IAM users, new access keys, and new Bedrock API keys, since the reported attackers created exactly those. - Review CloudTrail for bucket listings and secret searches from unfamiliar addresses. - Turn on threat detection. Our guide to AWS GuardDuty setup https://shattered.io/aws-guardduty-setup-organization-delegated-admin-12-steps/ walks through enabling it across an organization. - Set billing alarms on Bedrock and other model services so surprise usage gets noticed fast. The hours-long window in the Dragos result sets the pace. Rotation that happens next week protects little if the key was tried within hours of exposure. Automation, not a ticket queue, is the realistic answer. Five Predictions for AI Inference Security These are our forecasts based on the reported findings. None of them is a confirmed development. - Named disclosures will follow. Pressure from customers and press will likely push the affected providers, or Dragos, to publish more detail within months. - Retention claims will face audits. Expect enterprise buyers to demand third-party testing of Zero Data Retention promises before signing. - Secret scanning will move into the agent layer. Coding-agent vendors have a clear incentive to filter credentials before a prompt leaves the developer’s machine. - Reseller margins will compress. If trust becomes a buying criterion, the price advantage of marketplaces narrows. - Canary services for AI traffic will emerge. Security vendors will package bait-and-watch testing as a product for inference vendors and their customers. The common thread is measurement. Until now, customers took inference privacy on faith. This test shows how to measure it, and once measurement is cheap, faith gets harder to sell. For more coverage of model security, supply-chain risk, and the economics of AI services, see our AI and machine learning section https://shattered.io/ai-machine-learning/ . Frequently Asked Questions What did Dragos find about AI inference providers? Dragos sent about 2 million requests containing canary credentials and tracking data between July and November 2026. Reporting says at least three AI inference providers were compromised by attackers who copied incoming prompts, and some AWS keys were tested within hours. Which providers were compromised? The available reporting does not name them. The identities of the providers, resellers, the Chinese AI lab, and any router involved are not clearly established, so no company should be assumed guilty based on current public information. Did providers intentionally steal customer keys? That is unconfirmed. The documented claim is that providers were compromised or that logged data was exposed. Intent by any specific operator has not been established in the sources we reviewed. How many AWS keys were taken? Accounts disagree. One report says hundreds of AWS keys were harvested and another says roughly 200 live keys. Exact counts of stolen keys, API calls, and affected customers are inconsistent, so use the lower figure with care. What did attackers do with the stolen keys? Reports say they listed S3 buckets, searched AWS Secrets Manager https://shattered.io/aws-secrets-manager-pricing-vs-vault-key-vault-openbao-infisical-2026/ and Systems Manager Parameter Store, tried to enable Amazon Bedrock models, created Bedrock API keys, and added IAM users. Do Zero Data Retention settings protect me? Not reliably, based on the findings. Requests marked with Zero Data Retention or no data collection were not consistently handled according to those settings. Treat the flag as a request, and verify behavior independently when the data is sensitive. Can an AI router see my whole prompt? Yes, potentially. Reporting stresses that each intermediary handling an AI request can see the full prompt and response, which is why extra hops raise exposure. What is the fastest step to reduce risk today? Stop putting long-lived secrets into prompts. Scan agent context for credentials, use short-lived tokens with narrow permissions, and set alerts for unusual AWS activity such as new IAM users or Bedrock key creation.