Remote monitoring of SSD activity can reveal visited websites with up to 89% accuracy

Security researchers at Graz University of Technology (Austria) have described a side-channel attack that allows a malicious website to identify other sites and applications opened by a visitor by measuring SSD access latency using JavaScript in a standard browser sandbox.

The method, named FROST (Fingerprinting Remotely using OPFS-based SSD Timing), correctly identified visited sites with an accuracy of approximately 89%. Applications running on a test Mac were identified with an accuracy of approximately 96%. Such spying requires no other conditions apart from the victim visiting an infected site via a browser.

FROST utilises the Origin Private File System (OPFS) – a browser API that allows websites to create and store files on the user's local drive without requesting permission. known attacks on SSD side-channels required code execution via privileged kernel interfaces, but FROST does not require this.

The research team sent their findings to Google, Apple and Mozilla. Google stated that it does not consider the reading of such data to be a security vulnerability, Apple described the attack as currently being 'beyond the realm of possibility', and Mozilla acknowledged the findings but has not yet implemented any fixes.

The attack creates a large OPFS file on the SSD: browsers such as Chrome or Safari allow a website to occupy up to 60% of the total disk space. The file size must exceed the amount of available RAM so that every random 4 KB read hits the SSD rather than the OS cache. When an activity generates its own disk I/O, it creates measurable latency spikes in the attacker's read operations; these timing patterns are fed into a neural network trained to recognise specific websites and applications by their I/O signatures.

As everything happens at the data storage level, the attack works across different browsers. Launching the attacker's page in Chrome whilst the victim was viewing a page in Safari showed a throughput difference of just 3.38% compared to an attack in the same browser.

The full attack was only tested on an M2 Mac Mini with 8 GB of RAM and a 256 GB SSD. Regarding Linux, the researchers confirmed that they could measure SSD latency from the browser, but did not perform a full fingerprint classification, and Windows was not tested at all. The OPFS file must also be located on the same physical SSD as the monitored activity.

The biggest obstacle to carrying out the attack is the large file size: most users would notice that tens or hundreds of gigabytes had suddenly disappeared. Limiting the size of OPFS files to match the system's memory, or requiring explicit permission to create such files, could help prevent the attack.

Read also: Phishing instead of protection: fake Ledger Nano S+ crypto wallet sends PIN and seed phrases to the Internet