Skip to main content
Back
IntrusionAWS · Azure · SaaS
7 min read

LLMjacking of stolen cloud AI credentials through reverse proxies

Victim
Unnamed owners of compromised AWS, Azure and other AI service accounts
When
Not stated; proxies served DeepSeek after Dec 26, 2024Disclosed Feb 7, 2025
Entry
Credentials taken from vulnerable services such as Laravel and from software packages in public repositories, then tested with checker scripts against AI model services. The post does not say how the AWS and Azure keys in the proxies were taken.

Attackers took credentials from vulnerable services, such as Laravel, and from software packages in public repositories, tested them with scripts against AWS and Azure model services and loaded the valid keys into OpenAI Reverse Proxy instances that pool and resell them. After the DeepSeek models were released the proxies added them, and one proxy showed 55 DeepSeek keys. Sysdig estimates USD 49,595.83 of usage for one sold proxy and says its totals are minimums; it names no victim and gives no total of victims.

What happened

External Parties
Proxy operators
Proxy buyers
Sysdig
Internet
Cloud and AI Service APIswhere stolen keys authenticate
Vulnerable Services · Public Packages
Cloud AI Model Accounts
Vulnerable Service
Public Package
Bedrock Claude Access
Azure OpenAI Endpoint

Replaying the LLMjacking DeepSeek 2025 intrusion

  1. 1

    Stolen credentials: from services and packages

    Dates not stated

    Credentials are taken from vulnerable services, such as Laravel, and from software packages in public repositories; the post ties the AWS and Azure keys in the proxies to neither route.

  2. 2

    Keys checked against AWS and Azure model services

    Dates not stated

    Checker scripts test what each key reaches: AWS keys through STS and by billing, IAM and Claude access, Azure OpenAI endpoints by availability and content filtering status.

  3. 3

    Pooled behind proxies, resold and used

    After Dec 26, 2024; dates not stated

    Valid keys are loaded into open-source reverse proxies, DeepSeek keys are added, and buyers run models on the owners' accounts; Sysdig estimates almost USD 50,000 of usage for one proxy.

Impact

Stolen AI-service credentials, among them AWS and Azure keys, were pooled behind reverse proxies and used by buyers. For one sold proxy, Sysdig estimates almost USD 50,000 of usage over about 4.5 days, as one total that it does not split by provider. The number of victims is not stated, and Sysdig says the totals are minimums.

Outcome

Where cloud-provider controls fit

On AWS, a service control policy that refuses long-term access keys would have left a compromised service or package with role credentials that expire, and one that allows only approved Bedrock models would have limited a stolen key to the models on the list. Network conditions on each call would have limited the key to the organization's networks. Both bind only a member account of an AWS organization, and the post does not say whether the account behind the AWS key was one. On Azure, turning off public network access on the Azure AI Services account would have refused the proxies' requests even with a valid key. Turning off key access is documented to answer a request with an old key with HTTP 401, on a page that does not name Azure OpenAI. The vulnerable services, the public packages, the attackers' proxies and the OpenAI, Google AI, Mistral AI and DeepSeek accounts sit outside the cloud control plane: you own where your code keeps its secrets, and each vendor governs its own keys.

Could this have been prevented?

In hindsight, not as a whole. 1 control would have refused a step outright, and 10 controls would have limited one. 4 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.

1
Prevented
10
Contained
4
Not covered

How it could have been prevented

The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.

External Parties
Proxy operators
Proxy buyers
Sysdig
Internet
Cloud and AI Service APIswhere stolen keys authenticate
Vulnerable Services · Public Packages
Cloud AI Model Accounts
Vulnerable Service
Public Package
Bedrock Claude Access
Azure OpenAI Endpoint

Replaying the LLMjacking DeepSeek 2025 intrusion

1 · Stolen credentials: from services and packages

4 Contained2 Not covered4 on other clouds

2 · Keys checked against AWS and Azure model services

3 Contained2 on other clouds

3 · Pooled behind proxies, resold and used

1 Prevented3 Contained2 Not covered3 on other clouds

Key facts

  • Sysdig first discovered LLMjacking in May 2024 and published its DeepSeek research on February 7, 2025; it names no victim and no total of victims, and says the totals it observed are minimums
  • Attackers took credentials from vulnerable services, such as Laravel, and from software packages in public repositories; the post names no vulnerability, repository or package and does not say how the AWS and Azure keys in the proxies were taken
  • Checker scripts validated AWS keys with STS, sorted them by billing access, IAM profile update capability and access to Claude models, and tested Azure OpenAI endpoints for availability and content filtering status; the post names no API operation
  • DeepSeek-V3, released on December 26, 2024, was implemented in a proxy hosted on HuggingFace a few days later, and DeepSeek-R1, released on January 20, 2025, was implemented on January 21; one proxy showed 55 DeepSeek API keys
  • One sold proxy, with a 30-day token at USD 30, held 32 OpenAI keys, 13 OpenAI organizations, 2 Google AI keys, 1 AWS key and 1 Azure key, and served 132,651 prompts and 2.213 billion tokens to 18 active users in 395,555 seconds of uptime; Techzine reports that proxies are mainly shared through 4chan and Discord
  • Sysdig estimated USD 49,595.83 of usage for that proxy, almost USD 50,000 over about 4.5 days; its estimate for the aws-claude-opus line alone was USD 38,951.55 with Amazon Bedrock pricing, against USD 12,983.87 in the proxy's own figures, a gap the post does not explain

Sources

Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.

  1. 1.
    LLMjacking targets DeepSeekSysdig Threat Research Team · Feb 7, 2025 · ResearchCredential theft routes, checker scripts, reverse proxy design, DeepSeek adoption, proxy statistics, cost estimate, recommendations
  2. 2.
    LLM hijackers target DeepSeek APITechzine · Feb 10, 2025 · NewsReverse proxy sharing, DeepSeek key count

The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.

More cloud attack breakdowns

All breakdowns