Dark Reading is part of the Informa Tech Division of Informa PLC

This site is operated by a business or businesses owned by Informa PLC and all copyright resides with them.Informa PLC's registered office is 5 Howick Place, London SW1P 1WG. Registered in England and Wales. Number 8860726.

Vulnerabilities / Threats

09:00 AM
Connect Directly

Cerber Ransomware Now Evades Machine Learning

New variant has been broken into separate harmless-looking components to fool ML-based detection systems, Trend Micro says.

Cybercriminals have repeatedly shown an ability to innovate past whatever security controls organizations and industry have been able to throw in their way. So it is little surprise that some have begun taking a crack at machine learning tools.

Researchers at security vendor Trend Micro recently discovered a new version of the Cerber ransomware sample that appears designed specifically to evade detection by machine learning algorithms.

"The Cerber changes are really interesting as they're a direct response to changes in how some products are detecting malware," says Mark Nunnikhoven, vice president of cloud research for Trend Micro.

The newest version separates the different stages of the malware into multiple files and dynamically injects them into a running process, he says. "This helps to conceal them from various detection methods."

Like other ransomware threats, the new version of Cerber also is distributed via email. The email contains a link to self-extracting archive stored in a Dropbox account controlled by the attackers. The archive contains three files—one containing a Visual Basic script, the second a DLL, and the third, a binary file. The script is designed to load the DLL, which then reads the binary file and executes it. The binary file contains a new loader for Cerber and also the configuration settings for the malware.

The loader first checks to see if it is running in a sandbox or other protected environment. If it discerns that it's not in a protected environment, it injects the entire Cerber binary into one of several running processes, Trend Micro said in an alert this week.

"In their current form, some static machine learning-tools can have a hard time seeing the various pieces of the new configuration of Cerber," Nunnikhoven says. The malicious parts of it don't get analyzed, so the malware doesn't get flagged.

The reason is that static machine learning approaches look at the content of a file and evaluate the contents to see if they match malicious behaviors and attributes, he says.

But if the malicious content of the file is hidden for instance via encryption, or it is injected in real-time into a legitimate process, the content is not evaluated for suspicious behavior and attributes, he says.

"Say someone walks up to the door and they’ve got their hands behind their back. You look through the peephole and don’t see an immediate threat so you let them in," he says. You don't know until they are already in the house whether whatt they have in their hands is malicious or benign.

The latest innovations only make Cerber harder to detect via machine learning algorithms, he says. It can still be detected by other mechanisms. "The take-home message is that only using one technique to detect malware leaves you vulnerable if the criminals adapt to it."

News of Cerber’s latest tricks comes even as a new report from Carbon Black shows that many organizations remain unconvinced about the benefits of applying artificial intelligence and machine learning techniques to detect and stop cyber threats.

Nearly 75% of 410 security researchers that Carbon Black surveyed for the report describe AI-driven cybersecurity tools as being flawed, while 70% are convinced cyberattackers are capable of bypassing machine learning-based systems.

Mike Viscuso, co-founder and CTO of Carbon Black, says many current machine learning-based anti-malware tools are designed to stop attacks based on an inspection of files rather than behavior. They therefore miss the growing number of attacks that involve no malware files at all, he says.

Static, analysis-based approaches relying exclusively on files have been useful in the past. AI and ML-based tools can be useful in augmenting human decision-making and in spotting non-obvious relationships in massive volumes of security data. But they are of somewhat limited use in detecting non-malware attacks, he says.

Rather than using ML tools to look at individual files, organizations should be monitoring application and service activity, communications among processes, unauthorized requests to run applications, and changes to permission and credential levels, Vicuso says.

"If security tools are looking for just malware, they are missing an entire class of attacks that rely on native operating system tools to carry out nefarious actions. Attacks are evolving. So should [be] our defenses."

Related Content:


Jai Vijayan is a seasoned technology reporter with over 20 years of experience in IT trade journalism. He was most recently a Senior Editor at Computerworld, where he covered information security and data privacy issues for the publication. Over the course of his 20-year ... View Full Bio

Recommended Reading:

Comment  | 
Print  | 
More Insights
Threaded  |  Newest First  |  Oldest First
User Rank: Moderator
3/29/2017 | 5:21:50 PM
More FUD from legacy AV vendors about Machine Learning

First, let me be open and transparent by saying I work at Cylance as a pre-sales technical engineer. As a result, I have full access to our product and can test against malware. Here's my analysis and the results based on the samples Trend refers to in their post. The data below is unbiased, raw testing results.


f4dbbb2c4d83c2bbdf4faa4cf6b78780b01c2a2c59bc399e5b746567ce6367dd is a self extracting zip archive containing the other three files. Once it extracts, it automatically runs the vbs file it drops, 38oDr5.vbs. Cylance Protect's Script Control feature blocked the vbs file from ever executing. Threat neutralized pre-execution.


09ef4c6b8a297bf4cf161d4c12260ca58cc7b05eb4de6e728d55a4acd94606d4 is 38oDr5.vbs. For the sake of following the malware execution to see where else we'll block it, I completely disabled our Script Control feature, then tested again. When the vbs runs, it drops and tries to execute the malware dll, 8iqv.dll using the command:  "rundll32","8ivq.dll arzy949". We block and quarantine the 8iqv.dll malware sample pre-execution thanks to the industries most mature and advanced machine learning math model. Threat neutralized pre-execution.


e3e5d9f1bacc4f43af3fab28a905fa4559f98e4dadede376e199360d14b39153 is the 8ivq.dll our machine learning math model blocked pre-execution.


a61eb7c8d7a6bc9e3eb2b42e7038a0850c56e68f3fec0378b2738fe3632a7e4c has filename "x" without an extension. It's an obfuscated or crypted file with the malware's configuration parameters as well as a second stage malware component that gets loaded by the dll. It's the configuration file used by the malware. Since the malware was neutralized pre-execution at two different pre-execution points, there's no running dll malware to decrypt and run the content of X. The Cylance protected endpoint remained unscathed.


Now here's the kicker. The agent version I'm using in my lab hasn't been updated since November and the math model used to determine good from bad was created last summer. The results I describe were observed offline after reverting to my November snapshot and disabling network connectivity. The math is in the agent and it's future-proof. Now that's some real machine learning for you!

COVID-19: Latest Security News & Commentary
Dark Reading Staff 9/21/2020
Hacking Yourself: Marie Moe and Pacemaker Security
Gary McGraw Ph.D., Co-founder Berryville Institute of Machine Learning,  9/21/2020
Startup Aims to Map and Track All the IT and Security Things
Kelly Jackson Higgins, Executive Editor at Dark Reading,  9/22/2020
Register for Dark Reading Newsletters
White Papers
Current Issue
Special Report: Computing's New Normal
This special report examines how IT security organizations have adapted to the "new normal" of computing and what the long-term effects will be. Read it and get a unique set of perspectives on issues ranging from new threats & vulnerabilities as a result of remote working to how enterprise security strategy will be affected long term.
Flash Poll
How IT Security Organizations are Attacking the Cybersecurity Problem
How IT Security Organizations are Attacking the Cybersecurity Problem
The COVID-19 pandemic turned the world -- and enterprise computing -- on end. Here's a look at how cybersecurity teams are retrenching their defense strategies, rebuilding their teams, and selecting new technologies to stop the oncoming rise of online attacks.
Twitter Feed
Dark Reading - Bug Report
Bug Report
Enterprise Vulnerabilities
From DHS/US-CERT's National Vulnerability Database
PUBLISHED: 2020-09-25
The Telnet service of Rubetek RV-3406, RV-3409, and RV-3411 cameras (firmware versions v342, v339) can allow a remote attacker to gain access to RTSP and ONFIV services without authentication. Thus, the attacker can watch live streams from the camera, rotate the camera, change some settings (brightn...
PUBLISHED: 2020-09-25
A Cleartext Transmission issue was discovered on Rubetek RV-3406, RV-3409, and RV-3411 cameras (firmware versions v342, v339). Someone in the middle can intercept and modify the video data from the camera, which is transmitted in an unencrypted form. One can also modify responses from NTP and RTSP s...
PUBLISHED: 2020-09-25
The Telnet service of Rubetek cameras RV-3406, RV-3409, and RV-3411 cameras (firmware versions v342, v339) could allow an remote attacker to take full control of the device with a high-privileged account. The vulnerability exists because a system account has a default and static password. The Telnet...
PUBLISHED: 2020-09-25
Mitel MiCloud Management Portal before 6.1 SP5 could allow an attacker, by sending a crafted request, to view system information due to insufficient output sanitization.
PUBLISHED: 2020-09-25
Mitel MiCloud Management Portal before 6.1 SP5 could allow a remote attacker to conduct a SQL Injection attack and access user credentials due to improper input validation.