A Docker Hub experiment recorded the first use of an exposed Amazon Web Services (AWS) key just 7.86 seconds after researchers published the container image that held it. Across the experiment, the median wait for a first AWS request was 51.19 seconds, and all 100 test keys saw use within 15 minutes.

Researchers tracked 300 images across eight registries

Billy Tsouvalas and Nick Nikiforakis of Stony Brook University spent four months tracking what happened to credentials in public container images. They published 300 decoy images across eight registries. Alongside AWS, OpenAI and Anthropic keys, the images contained decoy credentials for databases and remote server access.

The researchers monitored requests made with the keys. They also set up alerts to tell them when someone built an image, ran a container or opened a shell. That gave them two ways to observe activity. They could see when a cloud service received a request and whether anyone had started the container that held the corresponding key.

The images drew more than 75,000 downloads. The AWS keys alone received over 140,000 API calls from about 5,000 distinct IP addresses. Some requests simply checked whether a key worked. Others listed resources available to the account, and in some cases the callers went on to use those resources. Whoever obtained the OpenAI and Anthropic keys spent the remaining balances on the test accounts.

Most AWS activity traced back to Docker Hub

Docker Hub accounted for nearly all the AWS activity the researchers observed. They also recorded some misuse of AWS keys from images on Harbor, another registry in the study. Despite the volume of requests, the monitoring system rarely detected anyone running a container. The researchers concluded that automated scanners generally found the credentials by inspecting image contents. They found no evidence that a single operator coordinated the scans.

The timing helps explain why the researchers want registries to check images for secrets before publication. In the fastest case, a publisher would have had less than eight seconds to spot the mistake and revoke the key before its first use. The team recommended warning publishers when a scan finds a likely credential. They also proposed limiting how frequently automated tools can access the public feed of new and updated images.

Tsouvalas said the team sent Docker its findings and full paper in May. Docker acknowledged the report. When the researchers ran a smaller test in August, they observed no change in the behavior they had documented.


0 responses to “Docker Hub test records cloud key use in under eight seconds”