The paradigm shift towards the cloud has dominated the technology landscape, providing organizations with stronger connectivity, efficiency, and scalability. As a result of ongoing cloud adoption, developers face increased pressures to rapidly create and deploy applications in support of their organization’s cloud transformation goals. Cloud applications, in essence, have become organizations’ crown jewels and developers are measured on how quickly they can build and deploy them. In light of this, developer teams are beginning to turn to AI-enabled tools like large language models (LLMs) to simplify and automate tasks.
Many developers are beginning to leverage LLMs to accelerate the application coding process, so they can meet deadlines more efficiently without the need for additional resources. However, cloud-native application development can pose significant security risks as developers are often dealing with exponentially more cloud assets across multiple execution environments. In fact, according to Palo Alto Networks’ State of Cloud-Native Security Report, 39% of respondents reported an increase in the number of breaches in their cloud environments, even after deploying multiple security tools to prevent them. At the same time, as revolutionary as LLM capabilities can be, these tools are still in their infancy and there are a number of limitations and issues that AI researchers have yet to conquer.
Risky business: LLM limitations and malicious uses
The scale of LLM limitations can range from slight issues to completely halting the process, and like any tool, it can be used for both helpful and malicious purposes. Here are a few risky characteristics of LLMs that developers need to keep in mind:
- Hallucination: LLMs may generate output that is not logically consistent with the input, even if the output sounds plausible to the user. The language model generates text that is not logically consistent with the input but still sounds plausible to a human reader.
- Bias: Most LLM applications rely on pre-trained models as creating a model from scratch is costly and resource-intensive. As a result, most models will be biased in certain aspects, which can result in skewed recommendations and content.
- Consistency: LLMs are probabilistic models that continue to predict the next word based on probability distributions – meaning that they may not always produce consistent or accurate results.
- Filter Bypass: LLM tools are typically built with security filters to prevent the models from generating unwanted content. However, these filters can be manipulated by using various techniques to change the inputs.
- Data Privacy: LLMs can only take encrypted inputs and generate unencrypted outlets. As a result, the outcome of a large data breach incident to proprietary LLM vendors can be catastrophic leading to effects such as account takeovers and leaked queries.
Additionally, because LLM tools are largely accessible to the public, they can be exploited by bad actors for nefarious purposes, such as supporting the spread of misinformation or being weaponized by bad actors to create sophisticated social engineering attacks. Organizations that rely on intellectual property are also at risk of being targeted by bad actors as they can use LLMs to generate content that closely resembles copyrighted materials. Even more alarming are the reports of cybercriminals using generative AI to write malicious code for ransomware attacks.
LLM use cases in cloud security
Luckily, LLMs can also be used for good and can play an extremely beneficial role in improving cloud security. For example, LLMs can automate threat detection and response by identifying potential threats hidden in large columns of data and user behavior patterns. Additionally, LLMs are being used to analyze communication patterns to prevent increasingly sophisticated social engineering attacks like phishing and pretexting. With advanced language understanding capabilities, LLMs can pick up on the subtle cues between legitimate and malicious communications.
As we know, when experiencing an attack, response time is everything. LLMs can also improve incident response communications by generating accurate and time reports to help security teams better understand the nature of the incidents. LLMs can also help organizations understand and maintain compliance with ever-changing security standards by analyzing and interpreting regulatory texts.
AI fuels cybersecurity innovation
Artificial intelligence will have a profound impact on the cybersecurity industry – and these capabilities aren’t strangers to Prisma Cloud. In fact, Prisma Cloud also provides the richest set of machine learning-based anomaly policies to help customers identify attacks in their cloud environments. At Palo Alto Networks, we have the largest and most robust data sets in the industry and we’re constantly leveraging them to revolutionize our products across network, cloud, and security operations. By recognizing the limitations and risks of generative AI, we will proceed with utmost caution and prioritize our customers’ security and privacy.
Author: Daniel Prizmant, Senior Principal Researcher at Palo Alto Networks
Daniel started his career developing hacks for video games and soon became a professional in the information security field. He is an expert in anything related to reverse engineering, vulnerability research, and the development of fuzzers and other research tools. To this day, Daniel is passionate about reverse engineering video games at his leisure. Daniel holds a Bachelor of Computer Science from Ben Gurion University.
