Typosquatting and credential stuffing are two of the most common ways that attackers are attempting to target companies' container infrastructure and the Docker-image supply chain, with attacks climbing nearly 600% in the second half of 2020 compared with the same period a year ago. That's according to a report released by cloud-native security provider Aqua Security on June 21.
Many attackers use passive scanning, utilizing services such as Shodan or tools such as Nmap to find servers hosting the Docker daemon or the Kubernetes container orchestration platform, attempting to attack those platforms using stolen credentials or vulnerabilities, according to the report. Another popular attack uses typosquatting — creating image names similar to legitimate images — and vanilla images that have a variant of a popular image, such as Alpine Linux, attempting to benefit from developer mistakes.
When attackers gain access, they most often install cryptominer software or attempt to escape the container and compromise the host system, says Assaf Morag, lead data analyst at Aqua Security.
"Attackers are constantly looking for new techniques to exploit containers and [Kubernetes]," he says. "They usually find an initial access to these environments and try escaping to the host and collect credentials, insert backdoors, and scan for more victims."
As companies move more of their infrastructure to the cloud, attackers have followed. A study of the publicly available images on Docker Hub conducted late last year found that 51% of the images had critical vulnerabilities and approximately 6,500 of the 4 million latest images — about 0.2% — could be considered malicious.
In addition, the developers who create and use containers often do not focus on security. A survey of 44 software images specifically used in neuroscience and medical data science found that a container built from the images had more than 320 different vulnerabilities on average.
The attackers know that misconfigurations are common and have used a variety of techniques to scan more frequently, the Aqua Security report states.
"This technique is very effective because the attackers use the infected hosts for the scanning operation, increasing the frequency of scanning activity and the chances of finding misconfigurations promptly," according to the report. "Some adversaries continue to use public search engines, such as Shodan or Censys, while others use scanning tools such as Masscan."
In one example of typosquatting, the company found an image by the name of "Tesnorflow," an attempt to profit from any misspellings of the well-known TensorFlow machine-learning package. Many data scientists use the container and may not pay attention to misspellings in the name, says Morag.
"If you are a data scientist who accidentally pulls and runs a Tesnorflow container image, you will also execute a cryptominer," he says. "We reported that to Docker Hub and they immediately removed it."
"It actually stole credentials and exfiltrated this data," the company says in the report. "We immediately informed this company and they have remediated this security issue."
Overall, attackers are using a greater number of images — an average of 3.78 per day — than the year before, and a greater number of attacks, 97 in the second half of 2020, up from 13 the same period the year before. A new honeypot is hit with its first attack within five hours, the report states.
Aqua Security's honeypots also detected attacks that attempt to use Kubernetes and automated build pipelines to build an application on a vulnerable server using images. Anti-defensive measures varied from the simple — such as packing executables to avoid signature scanners — to the more complex‚ such as disabling security measures and running code only in memory.
In 2021, attackers' focus appears to be shifting from compromising single containers and shifting to clusters of containers managed by Kubernetes, or K8s. The benefit for the attacker is expanding the scale of the eventual impact, Morag says.
"From the attacker’s perspective, the attack surface is bigger, securing K8s clusters is a bit more challenging, and if the attacker manages to find and attack a cluster, he has far more opportunities to gain from the attack," he says. "For instance, K8s allows attackers to execute multiple containers, so instead of a single cryptominer, he can run dozens and possibly come across more credentials and secrets."