Researchers from Stony Brook University have demonstrated that sensitive cloud credentials and API keys inadvertently included in public Docker Hub images are being harvested by automated scanners within seconds of being published. Billy Tsouvalas and Nick Nikiforakis conducted a four-month study involving 300 honeypot images across eight public container registries. By planting decoy credentials for AWS, OpenAI, and Anthropic, as well as SSH and database access keys, the researchers were able to measure the speed and scale of automated exploitation.
The findings, presented at the ACM Conference on Computer and Communications Security, reveal a highly efficient ecosystem of automated scanners. The researchers observed that the fastest instance of a leaked AWS key being utilized occurred just 7.86 seconds after the image was pushed to Docker Hub. The median response time for an initial API call was approximately 51 seconds, and every one of the 100 AWS keys deployed in the experiment was accessed within 15 minutes.
The scale of the activity was significant, with the honeypot images being pulled over 75,000 times. The AWS keys alone generated more than 140,000 API calls originating from 5,000 unique IP addresses. This traffic typically followed a predictable lifecycle: automated tools would first verify the validity of the credentials, then list available cloud resources, and finally proceed to consume them. For OpenAI and Anthropic keys, the objective was almost immediate exhaustion of the associated account balances.
The researchers concluded that these scanners, which they collectively refer to as ContainerBots, rarely execute the container images themselves. Instead, the attackers rely on static analysis, systematically unpacking image layers or querying registry APIs to extract filesystem contents. This method allows the scanners to bypass the need for container instantiation, making the credential theft process both rapid and resource-efficient.
While the activity was observed across multiple registries, Docker Hub accounted for the vast majority of the AWS key abuse. The researchers noted that there is no evidence suggesting these scanners are coordinated by a single entity; rather, the landscape appears to be populated by numerous independent, automated actors constantly monitoring public feeds for new uploads.
The researchers reported these findings to Docker in May, providing their full academic paper for review. According to the team, they received an initial acknowledgment but no further communication or indication that the platform had implemented defensive measures. A follow-up experiment conducted by the researchers in August confirmed that no changes had been made to the platform to mitigate the risk of secret exposure.
To address this vulnerability, the researchers suggest that container registries should implement automated scanning tools to detect and alert publishers to embedded secrets before an image is made public. Furthermore, they propose that registries consider restricting or rate-limiting access to public feeds of recently updated images, which would increase the difficulty for automated scanners to identify and target new uploads in real time.
The exposure of secrets in container images remains a persistent issue in software development, often requiring intensive remediation and credential rotation efforts once a breach is identified. The study highlights the critical need for developers to utilize secret management tools and automated pre-commit hooks to ensure that sensitive information is never committed to image layers in the first place.





