Detecting Obfuscated Command-lines with a Giant Language Mannequin

Within the safety trade, there’s a fixed, indisputable fact that practitioners should take care of: criminals are working extra time to consistently change the menace panorama to their benefit. Their strategies are many, they usually exit of their approach to keep away from detection and obfuscate their actions. In reality, one factor of obfuscation – command-line obfuscation – is the method of deliberately disguising command-lines, which hinders automated detection and seeks to cover the true intention of the adversary’s scripts.

Varieties of Obfuscation

There are just a few instruments publicly accessible on GitHub that give us a glimpse of what strategies are utilized by adversaries. One among such instruments is Invoke-Obfuscation, a PowerShell script that goals to assist defenders simulate obfuscated payloads. After analyzing a few of the examples in Invoke-Obfuscation, we recognized totally different ranges of the method:

Every of the colours within the picture represents a unique method, and whereas there are numerous kinds of obfuscation, they’re not altering the general performance of the command. Within the easiest kind, Gentle obfuscation modifications the case of the letters on the command line; and Medium generates a sequence of concatenated strings with added characters “`” and “^” that are typically ignored by the command line. Along with the earlier strategies, it’s doable to reorder the arguments on the command-line as seen on the Heavy instance, through the use of the {} syntax specify the order of execution. Lastly, the Extremely stage of obfuscation makes use of Base64 encoded instructions, and through the use of Base8*8 can keep away from a big quantity EDR detections.

Within the wild, that is what an un-obfuscated command-line would appear to be:

One of many easiest, and least noticeable strategies an adversary might use, is altering the case of the letters on the command-line, which is what the beforehand talked about ‘Gentle’ method demonstrated:

The insertion of characters which can be ignored by the command-line such because the ` (tick image) or ^ (caret image), which was beforehand talked about within the ‘Medium’ method, would appear to be this within the wild:

In our examples, the command silently installs software program from the web site The method used on this case is very stealthy, since it’s utilizing software program that’s benign by itself and already pre-installed on any laptop working the Home windows working system.

Don’t Ignore the Warning Indicators, Examine Obfuscated Components Shortly

The presence of obfuscation strategies on the command-line usually serves as a robust indication of suspicious (virtually all the time malicious) exercise. Whereas in some situation’s obfuscation could have a sound use-case, corresponding to utilizing credentials on the command-line (though this can be a very dangerous concept), menace actors use these strategies to cover their malicious intent.  The Gamarue and Raspberry Robin malware campaigns generally used this method to keep away from detection by conventional EDR merchandise. For this reason it’s important to detect obfuscation strategies as shortly as doable and act on them.

Utilizing Giant Language Fashions (LLMs) to detect obfuscation

We created an obfuscation detector utilizing giant language fashions as the answer to the consistently evolving state of obfuscation strategies. These fashions include two distinct components: the tokenizer and the language mannequin.

The tokenizer augments the command strains and transforms them right into a low-dimensional illustration with out shedding details about the underlying obfuscation method. In different phrases, the purpose of the tokenizer is to separate the sentence or command-line into smaller items which can be normalized, and the LLM can perceive.

The tokens into which the command-line is separated are basically a statistical illustration of widespread mixtures of characters. Subsequently, the widespread mixtures of letters get a “longer” token and the much less widespread ones are represented as separate characters.

Additionally it is vital to maintain the context of what tokens are generally seen collectively, within the English language these are phrases and the syllables they’re constructed from. This idea is represented by “##” on the planet of pure language processing (NLP), which suggests if a syllable or token is a continuation of a phrase we prepend “##”. One of the simplest ways to reveal that is to take a look at two examples; One among an English sentence that the widespread tokenizer gained’t have an issue with, and the second with a malicious command line.

For the reason that command-line has a unique construction than pure language it’s vital to coach a customized tokenizer mannequin for our use-case. Moreover, this tradition tokenizer goes to be considerably higher statistical illustration of the command-line and goes to be splitting the enter into for much longer (extra widespread) tokens.

For the second a part of the detection mannequin – the language mannequin – the Electra mannequin was chosen. This mannequin is tiny when in comparison with different generally used language fashions (~87% much less trainable parameters in comparison with BERT),  however remains to be capable of be taught the command line construction and detect beforehand unseen obfuscation strategies. The pre-training of the Electra mannequin is carried out on a number of benign command-line samples taken from telemetry, after which tokenized. Throughout this section, the mannequin learns the relationships between the tokens and their “regular” mixtures of tokens and their occurrences.

The subsequent step for this mannequin is to be taught to distinguish between obfuscated and un-obfuscated samples, which known as the fine-tuning section. Throughout this section we give the mannequin true constructive samples that had been collected internally. Nevertheless, there weren’t sufficient samples noticed within the wild, so we additionally created an artificial obfuscated dataset from benign command-line samples. Through the fine-tuning section, we give the Electra mannequin each malicious and benign samples. By displaying totally different samples, the mannequin learns the underlying method and notes that sure binaries have a better likelihood of being obfuscated than others.

The ensuing mannequin achieves spectacular outcomes having 99% precision and recall.

As we regarded via the outcomes of our LLM-based obfuscation detector, we discovered just a few new tips recognized malware corresponding to Raspberry Robin or Gamarue used. Raspberry Robin leveraged a closely obfuscated command-line utilizing wt.exe, that may solely be discovered on the Home windows 11 working system. However, Gamarue leveraged a brand new methodology of encoding utilizing unprintable characters. This was a uncommon method, not generally seen in studies or uncooked telemetries.

Raspberry Robin:


The Electra mannequin has helped us detect anticipated types of obfuscation, in addition to these new tips utilized by the Gamarue, Raspberry Robin, and different malware households. Together with the prevailing safety occasions from the Cisco XDR portfolio, the script will increase its detection constancy.


There are a lot of strategies on the market which can be utilized by adversaries to cover their intent and it’s only a matter of time earlier than we come upon one thing new. LLMs present new potentialities to detect obfuscation strategies that generalize nicely and enhance the accuracy of our detections within the XDR portfolio. Let’s keep vigilant and hold our networks secure utilizing the Cisco XDR portfolio.

We’d love to listen to what you suppose. Ask a Query, Remark Beneath, and Keep Linked with Cisco Safety on social!

Cisco Safety Social Channels



Leave a Reply

Your email address will not be published. Required fields are marked *