SMSMC Blog

Someone else is running AI on your cloud bill: stolen keys and AI services

Attackers use leaked cloud keys to run AI models on someone else's account and resell the access. How it has happened on AWS and Azure, and the controls that make a key less useful.

By Jim Venuto

Two cases, two clouds

In May 2024, Sysdig’s threat research team described an attack it named LLMjacking. The attackers took cloud credentials from a system running a vulnerable version of the Laravel framework (CVE-2021-3129), then used them to reach AI models hosted in the cloud, including Anthropic Claude models on Amazon Bedrock. Their tooling checked whether stolen keys worked against ten AI services, including AWS Bedrock, Azure and Google Cloud Vertex AI, and queried logging settings to avoid being noticed. Sysdig estimated that if the abuse went undiscovered, the victim could face over $46,000 a day in model usage at the prices and quotas of the time.

In February 2025, Microsoft’s Digital Crimes Unit named four people it alleges were part of a network it tracks as Storm-2139. According to Microsoft, members “exploited exposed customer credentials scraped from public sources” to get into customers’ Azure OpenAI accounts, changed what those services would do, and resold that access to other malicious actors. Microsoft had filed suit in December 2024 and amended its complaint to name the defendants.

Neither case needed a flaw in the AI service. Both used valid keys taken from somewhere the customer did not protect: a vulnerable server in one case, public sources in the other.

Why an AI key is a different kind of risk

Most small organizations that experiment with AI in the cloud start the same way: a developer or vendor creates a key, pastes it into an application or a script, and gets a prototype working. Nobody is assigned to watch it afterward.

Three things make that key worth stealing:

  • It turns directly into money. Model usage is billed per request, so a stolen key is a way to spend someone else’s budget, or to resell it.
  • It may not be watched. Teams that monitor sign-ins to their cloud console often do not monitor API calls to a model endpoint. Sysdig noted the attackers queried logging settings to avoid detection.
  • It carries your name. Whatever is generated through your account was generated by your account. In the Storm-2139 case, the access was resold to people the victims never met. That is a reputational and contractual problem as much as a billing one.

For a public body or a regulated business, there is a governance question too: if a key is used to send prompts and data to a model, who approved that use, and is it in the AI inventory?

Pointers that make a stolen key less useful

Azure and AWS each now have a way to take the long-lived key out of the picture. These are starting points; the right mix depends on how your applications are built.

  1. Azure: turn off key-based access where you can. Microsoft’s documentation for Foundry Tools (formerly Azure AI services) describes setting disableLocalAuth to true so that resources accept only Microsoft Entra authentication. There is a built-in Azure Policy that flags resources with key access still enabled. Microsoft notes the change can take from a few minutes to several hours to take effect, and suggests confirming that an old key now gets an HTTP 401 response.
  2. AWS: prefer short-term Bedrock API keys. AWS offers two kinds. Short-term keys use the IAM permissions of the identity that creates them and last up to 12 hours. Creating a long-term key creates an IAM user behind it. AWS recommends temporary credentials or short-term keys where security matters, and long-term keys with expiration dates for exploring Bedrock. Service control policies can cap how long a long-term key may live, or block the use of API keys altogether.
  3. Know what the logs do not show. AWS notes that Bedrock API calls are logged in CloudTrail, but the API keys, which travel in the request’s authorization header, are not. Pair logging with alerts on unusual model usage and on budget thresholds.
  4. Keep keys out of code and shared places. Both cases began with keys that were stolen or exposed. Store any remaining keys in a secrets manager, scan repositories for them, and rotate on a schedule and whenever someone with access leaves.
  5. Put AI resources in your AI inventory. Every model endpoint, who owns it, what data it is allowed to see and which applications call it.

Who does what

Leadership decides which AI services the organization may use and who may create access to them. Developers and vendors build applications to use managed identities or short-lived credentials, and disclose any long-lived keys. Your IT team or MSP applies the policies, sets up the alerts and budgets, and reviews the inventory on a schedule.

Where outside help fits

AI pilots often start outside the normal approval path, which is exactly how keys end up in the wrong places. SMC GRC.ai brings AI under the same governance as the rest of the business: a shadow AI inventory, AI vendor risk review, and an AI use policy that names owners. CyberIntelPro sets the architecture for how applications should authenticate to AI services across Azure, AWS and Google Cloud, and gives your team the change list. Privacy Medic covers the other half of the question, what personal data those AI services may receive.

Want to talk it through? Book a free 30-minute call with SMC GRC.ai.

Sources