Hacking AI Systems

After reading this chapter and completing the exercises, you will be able to do the following:

• Understand the different stages involved in an AI attack, including the steps from initial reconnaissance to the final impact.

• Identify and describe the different types of AI attack tactics and techniques used by attackers.

• Explain how attackers can develop resources and gain initial access to a system, including their methods for evading defenses and persisting within an environment.

• Evaluate the vulnerabilities of AI and ML models to unauthorized access and manipulation, as well as the potential impacts of such breaches.

• Illustrate how an AI attack is executed and how data is collected, staged, exfiltrated, and used for malicious intent.

• Design and implement proactive security measures to protect AI and ML systems from potential attacks.

• Understand AI attacks to develop response strategies to AI attacks, including incident handling, containment, eradication, and recovery.

Hacking FakeMedAI

The following is an attack against a fictitious company; however, it describes real-life attack tactics, techniques, and procedures (TTPs).

In the bustling tech hub of the Research Triangle area in North Carolina, a thriving AI startup named FakeMedAI created an innovative AI model that was revolutionizing the healthcare industry. Their proprietary model could predict the probability of a patient developing critical health conditions with remarkable accuracy. Unfortunately, FakeMedAI was about to face an adversary far more dangerous than any market competition.

Unbeknownst to FakeMedAI, their success had attracted the attention of a notorious Russian hacker group. The group started their operation by scouting FakeMedAI’s public digital footprint. They scrapped the company’s online resources, forums, press releases, and even LinkedIn profiles of key personnel to glean information about the architecture, usage, and potential vulnerabilities of the AI systems.

Attackers performed reconnaissance of publicly accessible sources, such as cloud storage, exposed services, and repositories for software or data, to find AI/ML assets. These assets might encompass the software suite employed to train and deploy models, the data used in training and testing, as well as model configurations and parameters. The attackers were especially interested in assets owned by or connected to the target organization because these are likely to reflect what the organization uses in a real-world setting. Attackers can locate these repositories of assets via other resources tied to the target organization, such as by searching their owned websites or publicly available research materials. These ML assets often grant adversaries insights into the ML tasks and methods used.

These AI/ML assets can boost an adversary’s efforts to construct a substitute ML model. If these assets contain parts of the real operational model, they can be employed directly to generate adversarial data. To obtain certain assets, registration might be necessary, which means providing details such as email/name and AWS keys, or submitting written requests, which might require the adversary to set up accounts.

Attackers gathered public datasets for utilization in their malicious activities. Datasets employed by the target organization, or datasets resembling those used by the target, could be of significant interest to attackers. These datasets can be stored in cloud storage or on websites owned by the victim. The datasets obtained aided the attackers in progressing their operations, planning attacks, and customizing attacks to suit the target organization.

Attackers also procured public models to utilize in their activities. They were interested in models that the target organization uses, or models that are analogous to those used by the target. These models might include model architectures, or pre-trained models that define both the architecture and model parameters, trained on a dataset. Attackers looked through different sources for common model architecture configuration file formats such as YAML or Python configuration files, and common model storage file formats such as ONNX (.onnx), HDF5 (.h5), Pickle (.pkl), PyTorch (.pth), or TensorFlow (.pb, .tflite). The models acquired were beneficial in propelling the attackers’ operations and are often used to customize attacks to match the victim’s model.

Having gathered a substantial amount of information, the hackers began crafting their strategy. They developed bespoke malware and set up a command and control (C2) server.

Using a carefully crafted phishing email disguised as an urgent message from the FakeMedAI CEO, the hackers targeted a low-ranking system administrator. The email contained a seemingly harmless PDF, which, once opened, installed the custom malware onto the system.

The company used a pre-release version of PyTorch, known as PyTorch-nightly. To their luck, the FakeMedAI system was breached. A harmful binary was uploaded to the Python Package Index (PyPI) code repository, compromising Linux packages. This malicious binary bore the same name as a PyTorch dependency, causing the PyPI package manager (pip) to install the harmful package instead of the genuine one.

This type of attack is a supply chain attack commonly referred to as a dependency confusion. It put at risk sensitive data on Linux machines that had installed the compromised versions of PyTorch-nightly via pip.

The malware propagated through the network, compromising credentials and escalating privileges until it reached the servers hosting the critical AI models. The hackers were cautious, avoiding high-traffic periods and masking their activities as regular network traffic to remain undetected.

Upon reaching the target servers, the malware initiated the main part of its program. It altered the AI model subtly, introducing a slight bias that would go unnoticed by regular integrity checks.

To maintain access to the system, the malware embedded itself in the boot records of the servers and established periodic communication with the C2 server. This allowed the attackers to monitor their progress and maintain control over the infected system.

To stay hidden, the malware used sophisticated evasion techniques like process hollowing and memory injection. It also removed event logs regularly to prevent the detection of its activities.

Adversaries were even able to craft adversarial data that hindered an AI model used for cybersecurity defensive operations from accurately recognizing the data’s contents. This technique was used to bypass a subsequent task where the attacker evaded detection.

While FakeMedAI remained oblivious, the hackers explored the compromised network to understand its topology and infrastructure better. They discovered additional data sets and resources that could be used in future attacks.

The hackers began collecting sensitive data, including patient records, proprietary ML algorithms, and internal communications, packaging them for extraction. The hackers transferred the collected data to their C2 servers. Using a slow and low technique, they made sure this process went unnoticed by the company’s intrusion detection systems.

Finally, the orchestrated attack was launched. The biased AI model began generating false predictions, causing chaos among healthcare providers and patients alike. The stolen data was sold on the dark web, and FakeMedAI’s reputation suffered a massive blow.

While this story is a fictional tale, it serves to illustrate the stages involved in a sophisticated AI system attack. It underscores the importance of a robust cybersecurity strategy that can prevent, detect, and respond to such intrusions.

On Chapter 4, you learned about the OWASP Top 10 for LLM Applications. We discussed threats such as prompt injection, insecure output handling, supply chain vulnerabilities, sensitive information disclosure, and others. Let’s explore some of the adversarial tactics and techniques against AI and ML systems.