Attackers are extracting and exploiting exposed cloud credentials from publicly accessible Docker images within seconds of their publication, according to a four-month experiment conducted by researchers at Stony Brook University. The fastest observed exploitation occurred just 7.86 seconds after an image containing a planted AWS key was uploaded to Docker Hub.
Researchers Billy Tsouvalas and Nick Nikiforakis investigated the activity by publishing 300 instrumented honeypot images across eight public container registries. The experiment examined how quickly automated scanners discover credentials embedded in container images and whether they subsequently attempt to use them.
Conducted over four months, the study involved planting AWS, OpenAI, and Anthropic API keys alongside decoy SSH credentials and database passwords. The researchers also deployed monitoring mechanisms designed to detect different interactions with their images, including image builds, container execution, and shell access.
The findings, presented at the 2026 ACM SIGSAC Conference on Computer and Communications Security (CCS), revealed extensive automated credential harvesting.
AWS Credentials Exploited Within Seconds
Across the experiment, the researchers recorded more than 75,000 image pulls. The planted AWS credentials alone attracted over 140,000 API requests originating from approximately 5,000 unique IP addresses.
The observed requests ranged from preliminary authentication checks to attempts to enumerate available cloud resources. Some interactions progressed further, indicating attempts to consume resources using the exposed credentials.
The earliest recorded AWS API request occurred just 7.86 seconds after an image was published on Docker Hub. The median time between publication and the first API request was 51.19 seconds. Furthermore, all 100 planted AWS keys had been accessed within 15 minutes of publication.
The OpenAI and Anthropic credentials experienced similar abuse. Automated activity rapidly consumed the balances associated with the researchers' test accounts, demonstrating that exposed AI service tokens can also become targets for immediate financial exploitation.
Docker Hub generated almost all observed AWS credential activity, although the researchers also recorded instances of abuse involving Harbor, another container registry included in the experiment.
Automated Scanners Extract Credentials Without Running Containers
The researchers collectively referred to the automated scanning activity as "ContainerBots." However, the designation does not establish that all observed scanners belonged to one threat actor or operated through a centrally coordinated infrastructure.
An important finding concerned how these bots accessed the exposed credentials.
Monitoring systems rarely detected attempts to execute the planted containers, suggesting that attackers generally did not need to run an image to retrieve its contents. Instead, they relied on static inspection techniques, including unpacking image layers and querying registry APIs to retrieve filesystem contents.
This approach allows scanners to examine files stored within container images without launching the applications themselves.
Container images package application code, dependencies, configuration files, and other resources into reusable layers. Consequently, credentials accidentally included during image creation can remain accessible through those layers even when developers do not intend to distribute them.
The researchers explained that their observations overwhelmingly involved static analysis rather than actual container execution.
Docker Hub Secret Exposure Is an Established Problem
The latest findings add to previous research documenting widespread credential exposure across public container repositories.
In December 2025, researchers at cybersecurity company Flare identified 10,456 Docker Hub images containing exposed secrets. These included approximately 4,000 AI-related API keys, alongside credentials capable of providing access to cloud environments, databases, Git repositories, and CI/CD infrastructure.
An earlier 2023 investigation involving 337,171 Docker images identified 52,107 valid private keys and 3,158 distinct API secrets across 28,621 images.
Such exposures create substantial remediation challenges. Removing an affected image or deleting a credential from its configuration does not invalidate copies already retrieved by automated scanners. Organizations must revoke compromised credentials, issue replacements, examine access logs, and assess whether unauthorized access occurred.
Researchers Recommend Additional Docker Hub Defenses
Tsouvalas and Nikiforakis recommend introducing automated secret detection before container images become publicly accessible. Docker Hub could scan uploaded images for embedded credentials and notify publishers when potentially sensitive information is discovered.
They also propose restricting or rate-limiting access to Docker Hub's
[…]
Content was trimmed to protect the source. Please visit the original article for the full text.
Read the original article:
Docker Hub Leaked Cloud Credentials Exploited Within 8 Seconds of Image Publication
