A cybersecurity illustration shows containerized systems, robotic agents, cloud services, and a glowing key beside a 7.86-second warning timer.
Eight seconds. That's how long it took, at best, for a stranger to find and use a cloud key that a developer had accidentally baked into a public Docker Hub image. Plenty of us take longer than that to find the Wi-Fi password. The finding comes from a controlled honeypot study by Stony Brook University researchers. It is a useful reminder for anyone who builds, pushes, or maintains container images, including Windows and Azure shops that publish to public registries.

A cybersecurity illustration shows containerized systems, robotic agents, cloud services, and a glowing key beside a 7.86-second warning timer. What the researchers did​

Billy Tsouvalas and Nick Nikiforakis presented the work at the ACM Conference on Computer and Communications Security (CCS). The paper is titled "Gone in 8 Seconds: Uncovering Malicious Bots Operating on Public Container Registries." Zenodo lists both authors as Stony Brook University researchers. The authors also published an artifact bundle on Zenodo.

According to BleepingComputer's report, the team published 300 honeypot images across eight public container registries over four months. The images carried planted AWS, OpenAI, and Anthropic API keys, plus decoy SSH and database credentials. The researchers also added detection tooling. It would alert them if an image was built, run, or had a shell opened in it.

The report gives these headline numbers:

  • The images were pulled more than 75,000 times.
  • The planted AWS keys received over 140,000 API calls from 5,000 unique IP addresses.
  • The AWS traffic ranged from checking whether a key worked, to listing resources, to sometimes consuming them.

The report doesn't give the exact dates of the experiment. It also doesn't say how the 300 images were split across registries, or how many keys of each type were planted. I'm not going to guess at those.

The timing numbers, read carefully​

The headline says "within seconds," so it's worth separating the three timing figures. They measure different things.

MeasureResult
Fastest observed use of an AWS key7.86 seconds after publication on Docker Hub
Median time to first API call51.19 seconds
All 100 planted AWS keys usedwithin 15 minutes

So 7.86 seconds is the best case for the attacker, not the typical case. The median is still under a minute, and every AWS key was touched within 15 minutes. If you're counting on a human noticing your mistake and fixing it in time, those numbers say you won't have the time.

AI keys went the same way. The exposed OpenAI and Anthropic keys were quickly hijacked to exhaust the test-account balances. The report gives no timings, call counts, or dollar figures for those services.

Nobody had to run the container​

This may be the most operationally important finding. Monitoring rarely detected anyone actually running the honeypot images, so the authors concluded that scanners generally pulled credentials by inspecting image layers directly. The researchers said the scanners work through static analysis. That means unpacking layers or querying the registry API to read filesystem contents, without ever starting the container.

In practice, a secret doesn't have to be reachable from a running service to be at risk. If it sits in any layer of a public image, it can be read. This is a design-level problem with how images are built. A common example is a COPY of a config file, or a build argument that ends up in a layer. A later layer that deletes the file doesn't help, because the earlier layer still holds it. That last point is general container knowledge, not a finding of this study.

"ContainerBots" is not one gang​

The researchers coined "ContainerBots" for the automated scanners that touched their images. BleepingComputer stresses that it is a loose umbrella term, and the scans showed no evidence of being coordinated by the same operators. So this is not evidence of a single botnet or a named threat actor. The 5,000 IP addresses are also a count of addresses seen, not of distinct people or groups.

Which registries were affected​

Docker Hub accounted for nearly all AWS activity, while Harbor, one of the other registries in the experiment, also saw some AWS key abuse. The report gives no fuller per-registry breakdown. That means this study can't tell you that other registries are safe. It only says that most of the observed AWS abuse happened on Docker Hub.

What the researchers want Docker to do​

The researchers made two recommendations:

  • Scan images for secrets before they go live, and warn publishers when an image appears to contain any.
  • Restrict or rate-limit access to Docker Hub's public feed of recently published or updated images, to hinder mass scanning.

Tsouvalas told BleepingComputer that the team reported the findings and the full paper to Docker in May. They got an acknowledgment but no follow-up. A small test in August found no changes. This is the researchers' account, not an independent check of Docker's current controls. I haven't seen a Docker response on the record.

What to do about it​

The study doesn't supply a remediation checklist. The following is general practice that follows from its findings, not something the researchers tested:

  1. Treat any secret in a public image as compromised. Deleting or replacing the image doesn't revoke the credential. Rotate or revoke the key, then review its activity logs. BleepingComputer notes that revoking leaked keys takes significant work and coordination, which is a good reason to practice before an incident.
  2. Keep secrets out of image contents. Inject them at runtime from a secrets manager or the orchestrator, not at build time.
  3. Scan before you push. Add secret scanning to CI so a bad image fails the pipeline instead of reaching the registry.
  4. Use short-lived, narrowly scoped credentials. If a key must leak, the damage depends on what it can do and how long it works.
  5. Set billing and usage alerts. The OpenAI and Anthropic results show that the cost of a leaked AI key can be a drained balance.

How common is this?​

Earlier research shows this isn't a new problem. A 2023 RWTH Aachen study, "Secrets Revealed in Container Images," is a separate dataset and method from this one, so its numbers shouldn't be mixed with the honeypot results. That study analyzed 337,171 Docker Hub images and 8,076 images from other registries. It found that 8.5% included secrets, among them 52,107 private keys and 3,158 API secrets. Together, the two studies suggest that secrets in images are common and that attackers pick them up very quickly. The honeypot study can't say how many real public images contain secrets.

Bottom line​

This is a controlled experiment, so treat the numbers as what the researchers observed with planted credentials. The practical conclusion still holds. For any credential that reaches a public registry, assume the clock started at publication, and that you may already be late. For teams building container images on Windows, in Azure pipelines, or anywhere else, the cheapest fix is to stop secrets from entering the image in the first place.

 

References

  1. Cloud keys leaked in Docker Hub images are stolen within seconds BleepingComputer 2026-11-17T16:27:17-05:00
  2. Gone in 8 Seconds: Uncovering the Malicious Bots Operating on Public Container Registries zenodo.org
  3. Secrets Revealed in Container Images: An Internet-wide Study on Occurrence and Impact | COMSYS | RWTH Aachen University comsys.rwth-aachen.de