The artificial intelligence (AI) stack built into Windows can act as a channel for malware transmission, a recent study has demonstrated.
Using AI in malware
Security researcher hxr1 discovered a far more conventional method of weaponizing rampant AI in a year when ingenious and sophisticated quick injection tactics have been proliferating. He detailed a living-off-the-land attack (LotL) that utilizes trusted files from the Open Neural Network Exchange (ONNX) to bypass security engines in a proof-of-concept (PoC) provided exclusively to Dark Reading.
Impact on Windows
Programs for cybersecurity are only as successful as their designers make them. Because these are known signs of suspicious activity, they may detect excessive amounts of data exfiltrating from a network or a foreign.exe file that launches. However, if malware appears on a system in a way they are unfamiliar with, they are unlikely to be aware of it.
That's the reason AI is so difficult. New software, procedures, and systems that incorporate AI capabilities create new, invisible channels for the spread of cyberattacks.
Why AI in malware is a problem
The Windows operating system has been gradually including features since 2018 that enable apps to carry out AI inference locally without requiring a connection to a cloud service. Inbuilt AI is used by Windows Hello, Photos, and Office programs to carry out object identification, facial recognition, and productivity tasks, respectively. They accomplish this by making a call to the Windows Machine Learning (ML) application programming interface (API), which loads ML models as ONNX files.
ONNX files are automatically trusted by Windows and security software. Why wouldn't they? Although malware can be found in EXEs, PDFs, and other formats, no threat actors in the wild have yet to show that they plan to or are capable of using neural networks as weapons. However, there are a lot of ways to make it feasible.
Attack tactic
Planting a malicious payload in the metadata of a neural network is a simple way to infect it. The compromise would be that this virus would remain in simple text, making it much simpler for a security tool to unintentionally detect it.
Piecemeal malware embedding among the model's named nodes, inputs, and outputs would be more challenging but more covert. Alternatively, an attacker may utilize sophisticated steganography to hide a payload inside the neural network's own weights.
As long as you have a loader close by that can call the necessary Windows APIs to unpack it, reassemble it in memory, and run it, all three approaches will function. Additionally, both approaches are very covert. Trying to reconstruct a fragmented payload from a neural network would be like trying to reconstruct a needle from bits of it spread through a haystack.
