This paper presents a new attack allowing attackers to exfiltrate data from isolated, air-gapped computers via Wi-Fi frequencies. We show that malware in a compromised air-gapped computer can generate signals in the Wi-Fi frequency bands. The signals are generated through the memory buses - no special hardware is required. Sensitive data can be modulated and secretly exfiltrated on top of the signals. We show that nearby Wi-Fi-capable devices (e.g., smartphones, laptops, and IoT devices) can intercept these signals, decode them, and send them to the attacker over the Internet. We utilized the physical layer information exposed by the Wi-Fi chips to extract the signals. We further implemented the transmitter and receiver and discussed design considerations and implementation details. We evaluated this covert channel in terms of bandwidth and distance and presented a set of countermeasures. Our evaluation shows that data can be exfiltrated from air-gapped computers to nearby Wi-Fi receivers located meters away at bit rates of 16 b/sec.

In this paper, we introduce a new covert channel that enables attackers to leak data acoustically from highly secure, air-gapped and audio-gapped systems. The proposed malware can exploit the computer power supply unit (PSU) to play sounds and use it as an out-of-band, secondary speaker with limited capabilities. Our method enables the production of audio streams in a frequency band of 0-24kHz and playing audio files (e.g., WAV) from a computer power supply without the need for audio hardware or speakers. Furthermore, we developed techniques that exploit the multiple CPU cores and their corresponding switching frequencies. It enables the production of multi-channel soundtracks and controlling the bit-depth and volume of sound played through the power supply. Sensitive information (documents, biometric data, encryption keys, etc.) can be covertly transmitted to a nearby receiver (e.g., smartphone). We show that our technique works with various types of systems: PC workstations and servers, as well as embedded systems and IoT devices that have no audio hardware at all. We also show that the attack can operate from an ordinary user mode process, run on virtual machines monitors (VMMs), an