An Austrian research team recently disclosed a new browser side-channel attack method. It only needs to rely on weak signals from the code running in the web page and the device storage hardware to "snoop" on the user's online behavior without relying on traditional cookies, click tracking scripts or common fingerprint recognition technology. This approach takes advantage of the timing characteristics of solid-state drives (SSDs) when processing data requests.
This technology is named FROST, which stands for "OPFS-based SSD timing remote fingerprinting" (fingerprinting remotely using OPFS-based SSD timing). The core idea is that when different processes compete for storage access, they will leave subtle but observable differences in read and write latencies. By monitoring these latency changes, researchers can infer which websites or applications are running on the device.
From a technical perspective, FROST is a typical side-channel attack, that is, it does not directly read sensitive data, but infers information through "side-channel signals" of system behavior. In this case, the "side channel signal" is the access delay of the SSD: when multiple programs perform read and write operations at the same time, the response time of the SSD will change accordingly, and these changes can be captured and analyzed inside the browser.
What's special about FROST is that it runs entirely within a browser environment and requires no additional software to be installed locally. The attacker only interacts with the Origin Private File System (OPFS) through JavaScript - OPFS was originally designed to provide isolated local storage space for websites. However, although sandbox isolation is implemented at the software level, the underlying hardware is still shared, which is the key to information leakage.
In actual operation, the attack script will first create a large file in OPFS, then continuously perform random reads on the file, and accurately record the time spent on each read operation. If other applications or browser tabs in the system are frequently accessing the same SSD at this time, the latency of these read operations will fluctuate measurably, and the long-term accumulation of data will form a discernible time series pattern.
An attacker can obtain these latency traces by continuously measuring SSD contention and then train and classify them using a convolutional neural network, the researchers said in their paper. Once the model is trained, it can "fingerprint" user activity based on the new latency traces, i.e. determine what actions are taking place on the host system, such as whether a specific website is opened or a certain type of application is run.
Notably, the team found that this approach works equally well across browsers, suggesting that exploitable signals come more from underlying system behavior than implementation differences in a particular browser. It also highlights a broader trend: As browsers evolve into platforms capable of running complex applications, they increasingly interact with underlying system resources, increasing the surface for unintended data exposure.
Researchers pointed out that today's browsers have long been upgraded from "document viewing tools" to complex platforms that can host complete office suites, image and video editors, and even integrated development environments (IDEs) provided by companies such as Google, Microsoft, and Adobe, all running within the browser. Although these enhancements have expanded the capabilities of web applications and created new usage scenarios, they have also significantly expanded the browser's attack surface, and some features have been proven to introduce new security vulnerabilities.
In terms of practical feasibility, FROST still has certain limitations. For example, the attack needs to generate large files of at least 1GB level in OPFS, which may attract user attention or trigger storage space reminders; at the same time, this method relies on the monitored activity and the attack script running on the same physical SSD. If the target application is located on another hard disk, the signal may be weak and difficult to detect.
In terms of the experimental platform, the research team demonstrated the complete attack process on a system equipped with an Apple M2 chip, verified the core mechanism in a Linux environment, and found that similar behavior also existed. Hannes Weissteiner, co-author of the paper, said that the underlying technology performs similarly on macOS and Linux, so the complete classification method is expected to perform similarly on these two systems; in theory, as long as a certain type of system activity stably produces a specific SSD access pattern, the model can be trained to identify it, but the team has not yet tested it on Windows systems.
There is currently no evidence that this technology has been used maliciously outside of a research environment, but it is further evidence that the code running in the browser can "sense" system behavior deeper than traditional knowledge. The researchers suggest that browser vendors can reduce risks by imposing stricter restrictions on OPFS, such as setting file size limits or monitoring for abnormal storage access patterns to block suspicious behavior in a timely manner.
According to reports, relevant research results are expected to be officially published at the DIMVA conference in July. As more similar research advances, how to control underlying side channel risks while ensuring rich functionality of web applications is likely to become a long-term issue in the field of browser security.