
Two Stony Brook University researchers, Billy Tsouvalas and Nick Nikiforakis, measured how fast attackers grab credentials that developers accidentally leave inside public container images. Over four months they published 300 honeypot images across eight public registries, embedding AWS, OpenAI and Anthropic API keys along with decoy SSH and database credentials. They presented the results at the ACM Conference on Computer and Communications Security (CCS), as reported by BleepingComputer.
The numbers are short. The fastest use of a planted AWS key came 7.86 seconds after an image was published on Docker Hub; the median time to the first API call was 51.19 seconds, and all 100 planted AWS keys were used within 15 minutes. The images were pulled more than 75,000 times, and the AWS keys received over 140,000 API calls from about 5,000 IP addresses, ranging from checking whether a key works to listing resources and sometimes consuming them. The exposed OpenAI and Anthropic keys were quickly used to drain the test accounts’ balances. Docker Hub accounted for nearly all AWS activity, with some abuse also seen on Harbor.
Monitoring rarely detected anyone actually running the honeypot images. The researchers concluded that the scanners they call “ContainerBots” mostly unpack image layers or query the registry API and read the files directly, without starting the container. The scans showed no evidence of a single coordinated operator. The researchers recommend that Docker Hub scan images for secrets before they go live and warn publishers, and that it restrict or rate-limit the public feed of recently published or updated images. Tsouvalas told BleepingComputer they reported the findings and the paper to Docker in May, got an acknowledgment but no follow-up, and saw no change in an August re-test.
How to check if you’re affected
Affected products are any container images published to a public registry, Docker Hub in particular, that ever contained an API key, cloud credential, SSH key or database password in a layer. The study does not name real victims; it shows how quickly any such secret is likely to be found.
The study’s own recommendations are for registry operators. The following are our suggestions, not from the researchers:
- Review the images your team has pushed publicly and check whether any layer, build argument or environment file carried a credential.
- Treat any key found that way as already used by someone else: revoke and reissue it rather than only deleting the image. BleepingComputer notes that revoking leaked keys takes significant work and coordination.
- Check your cloud provider’s access logs for calls from addresses you do not recognise around the publication time.
