1. Basic Information
- Article Title: Cloud keys leaked in Docker Hub images are stolen within seconds
- Publisher: BleepingComputer
- Publication Date: 2026-11-17
- Collection Date: 2026-11-18T08:00:28+09:00
- Original Source: BleepingComputer
- Related Sources: ContainerBots project, Author Publication List
- Severity (Author Priority): High
- Basis for Severity: In an experiment with 300 honeypot images, AWS keys were used in as little as 7.86 seconds (median 51.19 seconds), and all 100 keys were utilized within 15 minutes.
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
- Prepare decoy credentials for AWS, OpenAI, Anthropic, SSH, and databases.
- Publish 300 honeypot images across eight registries over a four-month period.
- Measure pull, build, run, shell, and API key usage activities.
- 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.
-
ContainerBotsis 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.
Top comments (0)