Decoding the world of cybersecurity

Container credentials exposed to exploitation in seconds

Researchers found that cloud credentials embedded in public container images could be discovered and used in seconds, with automated scanners exploiting deliberately exposed test keys across multiple registries.

Container credentials exposed to exploitation in seconds
Summary
  • Stony Brook University researchers deployed 300 instrumented container images across eight public registries.
  • The fastest observed AWS credential use occurred 7.86 seconds after publication, with a median of approximately 51 seconds.
  • Automated systems generated extensive API activity, demonstrating the limits of detecting accidental secret exposure only after software publication.

Cloud credentials accidentally included in public container images can be discovered and used almost immediately, according to research demonstrating that automated systems exploited a deliberately exposed test key less than eight seconds after publication.

Researchers at Stony Brook University deployed 300 instrumented container images across eight public registries over four months, monitoring activity involving embedded credentials and other decoy information.

The fastest observed use of an Amazon Web Services credential occurred 7.86 seconds after publication. The median time to the first observed use was approximately 51 seconds, while every AWS key in the initial group of 100 was accessed within 15 minutes.

The research, titled Gone in 8 Seconds: Uncovering the Malicious Bots Operating on Public Container Registries, was presented at the 2026 ACM Conference on Computer and Communications Security.

Automated scanners inspect published images

Researchers Billy Tsouvalas and Nick Nikiforakis designed a controlled experiment using container images containing monitored credentials, including AWS keys and API credentials associated with AI services.

Container images package applications with supporting software and configuration information, allowing developers to distribute systems that can run consistently across different computing environments.

They are also capable of retaining sensitive information unintentionally introduced during development or automated builds. Credentials may appear in configuration files, environment settings, or intermediate filesystem layers created during construction of the image.

Removing a secret from the final visible filesystem does not necessarily erase it from earlier layers. Where those layers remain part of the published image, the information may still be recoverable by anyone able to retrieve the artefact.

Automated scanning systems can examine those files without compromising the organisation that created the image or executing its application.

In the experiment, researchers observed credential use shortly after publication, indicating that some systems were monitoring registries for new artefacts and inspecting them at considerable speed.

The median observed delay of approximately 51 seconds shows that rapid activity was not limited to one isolated event, although the experiment does not establish identical behaviour across every public registry or type of credential.

The researchers also recorded substantial differences between individual observations, reflecting variation in how quickly systems discovered the images and attempted to use the exposed keys.

Exposed credentials triggered extensive API activity

Across the study, the researchers recorded more than 140,000 AWS API calls originating from over 5,000 distinct IP addresses.

They also observed rapid consumption of the available balances associated with monitored OpenAI and Claude API credentials. These findings indicate that some scanning activity progressed beyond collecting information to attempting to obtain usable computing or AI services.

The experiment used controlled credentials and decoy infrastructure. Its results do not mean that 140,000 production systems were compromised or that every observed request represented a separate attacker.

Nevertheless, the activity demonstrates how exposed credentials can become valuable to automated systems even when the legitimate software application containing them is never deployed.

An attacker possessing a cloud credential may be able to access resources or create charges depending on the permissions attached to the identity. An API key can likewise permit consumption of paid services if adequate limits and controls are absent.

Credentials with narrow permissions create different risks from those providing administrative access, but even restricted keys may reveal information or support unauthorised operations within their permitted scope.

Pre-publication checks become increasingly important

Traditional approaches to secret exposure often depend on detecting a published credential and then revoking it. The Stony Brook findings show that automated use may begin before a human administrator or scheduled security scan identifies the mistake.

Removing an image from a public registry cannot undo requests already made with an exposed key. Nor does deletion guarantee that previously downloaded copies have disappeared.

Where exposure occurs, response involves revoking affected credentials, reviewing their permissions, and examining service activity for unauthorised access or spending.

Preventive controls consequently need to operate during development and before publication. These may include scanning source repositories and container layers, using secret-management services, and preventing credentials from being embedded in build artefacts.

Development teams also need to understand how container layers preserve information. A build process that copies credentials into an intermediate layer and later deletes them may still create an image containing recoverable secrets.

Automated deployment pipelines require particular attention because their credentials may grant access to registries, source repositories, or production cloud resources. Such permissions should be limited to the operations required and revoked when no longer necessary.

The experiment does not establish that every exposed credential will be used within eight seconds. It provides measured evidence that automated harvesting can operate on a timescale much shorter than conventional reactive security processes.

With the earliest observed exploitation occurring in 7.86 seconds and the median close to 51 seconds, the researchers’ findings place greater emphasis on preventing secret exposure than relying exclusively on discovery and correction after a container becomes publicly accessible.

×