Dev.to Security 🔐 Cybersecurity 👁 0 📖 3 min read

ContainerBots: AWS Keys in Public Container Images Used as Little as 7.86 Seconds After Publication

1. Basic Information Article Title: Gone in 8 Seconds: Uncovering Malicious Bots Operating on Public Container Registries Publisher: Stony Brook University Publication Date: 2026-11-17 Collection Date: 2026-11-18

1. Basic Information

2. Key Takeaways

BleepingComputer reports that researchers published images containing decoy secrets across eight public container registries to measure the harvesting and usage of AWS and LLM API keys. The experimental conditions and figures below are based on this report and have not been verified against the primary source.

3. Research Objectives

  • Measure the speed, scale, and methods of automated harvesting of secrets accidentally exposed in public registries.

4. Methodology and Workflow

  1. Prepare decoy credentials for AWS, OpenAI, Anthropic, SSH, and databases.
  2. Publish 300 honeypot images across eight registries over a four-month period.
  3. Measure pull, build, run, shell, and API key usage activities.
  4. Analyze source IPs, API actions, time to first use, and container execution status.

5. Inputs and Outputs

  • Input: Decoy credentials embedded in public image layers.
  • Output: Observation data covering over 75,000 pulls, more than 140,000 AWS API calls made using the decoy keys, and approximately 5,000 IP addresses.

6. Evaluation Design

  • Controlled credentials rather than actual customer secrets were used to measure external scanner responses by registry and time.

7. Results

  • The fastest AWS key usage occurred 7.86 seconds after publication to Docker Hub, with a median time to first call of 51.19 seconds.
  • All 100 AWS keys were utilized within 15 minutes.
  • Scanners frequently read the file system via layer extraction or registry API without starting containers.
  • OpenAI and Anthropic test balances were also consumed rapidly.

8. Practical Implications

  • Deleting secrets from public registries after pushing is insufficient. Organizations must implement CI pre-push secret scanning and automate immediate revocation upon leakage.

9. Prerequisites, Limitations, and Application

  • The prerequisite is that a layer containing secrets is published to a public registry. Deleting files in subsequent layers may still leave them in past layers.
  • ContainerBots is not a single threat actor name, but a research designation grouping multiple scanners.

10. Facts / Inference / Hypothesis

Facts

  • A controlled experiment spanning 300 images, 8 registries, and 4 months.
  • Docker Hub accounted for nearly all AWS key usage, with some exploitation also confirmed on Harbor.
  • Researchers reported their findings to Docker in May 2026 and stated that they did not observe any defensive changes in a small-scale re-test in August.

Inference

  • Response SLAs for credential leaks should be designed assuming pre-publication blocking and revocation within seconds rather than minute-level response.

Hypothesis

No additional hypotheses. Unverified items are listed in "Open Questions and Further Research".

11. MITRE ATT&CK Mapping

ID Technique Confidence Basis
T1552.001 Unsecured Credentials: Credentials In Files medium Mapping based on reports of credential harvesting from files within image layers. Unverified in the primary source. Definition
T1580 Cloud Infrastructure Discovery medium Mapping based on reports of AWS resource enumeration. Unverified in the primary source. Report

12. Open Questions and Further Research

  • Experimental conditions, numerical figures, registry distribution, and details of the report to Docker and subsequent re-testing remain unverified in the primary source.
  • The identity of each scanner, relationships among operators, the number of real-world victim incidents, and future preventive measures by registries remain undetermined.

13. Implications for Defenders

Require secret scanning of container images, including all underlying layers, in build pipelines, and monitor by correlating public registry events, initial cloud API usage, and LLM API cost anomalies.

14. Summary by Target Audience

  • For SOCs: Correlate registry publication, secret detection, cloud API usage, and LLM utilization on a second-by-second basis.
  • For Administrators: Avoid including secrets in build args or layers; implement pre-push scanning and automated revocation.
  • For Users: This study uses decoy credentials and does not indicate actual key leaks from specific companies.
📰 Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.