Security researchers at Austria's Graz University of Technology have uncovered vulnerabilities in file notification systems that leak sensitive user activity data across Android, Linux, macOS, and Windows. The flaws affect systems dating back decades—Linux's inotify since 2005, Android's FileObserver since 2008, Windows' ReadDirectoryChangesW since 2000, and macOS's FSEvents since 2007. These subsystems notify applications when files are opened, modified, written, or deleted, and while they don't expose file contents, the event data creates a side channel that allows malicious users to infer what other users are doing on the same machine.
The vulnerabilities enable multiple attack vectors with alarming accuracy rates, according to findings detailed in a paper scheduled for presentation at ACM CCS 2026 this November in The Hague. On Linux, monitoring the readable directory /dev/input produces notifications on every keystroke, which researchers converted into a local inter-keystroke timing attack achieving 93.1 to 100 percent keystroke accuracy across seven users and 100 percent accuracy for remote SSH sessions. On Windows, watching the root directory C:\ exposes the full path of every file accessed anywhere on the system, regardless of permissions, enabling real-time tracking of Firefox website visits with 97.8 percent accuracy. Android's FileObserver bypasses the FUSE isolation layer intended to protect per-app storage, allowing a permissionless application to monitor WhatsApp's private folder and observe exactly when photos, videos, and documents are sent, received, or deleted based on filename and timestamp. The research team also achieved 87.9 percent accuracy in website fingerprinting on the top 100 sites and successfully executed an authentication-prompt redress attack on KDE Plasma 6 under Wayland.
The core issue is that unprivileged users can access file notification subsystems, and on Linux and Windows, file information remains available even without read permissions, according to doctoral student Sudheendra Raghav Neela and colleagues. Linux partially addressed its vulnerability—CVE-2025-68788—in December 2025 through patches in kernels 5.10.248, 5.15.198, 6.1.160, 6.6.120, 6.12.64, and 6.18.3 that prevent access and modify events on special files in /dev/. Despite responsible disclosure to security teams between August and October 2025, Android has received no mitigation. Apple's macOS proved least vulnerable because no bypasses were discovered for reading private directories, though FSEvents still allows monitoring of various system changes visible in .plist files, including audio settings, power configurations, Bluetooth devices, network DNS changes, and application installations. Microsoft told researchers that Windows' behavior is "by-design" and an undocumented feature—a response nominated for the lamest vendor response at the Pwnie Awards 2026.
The research team argues that all major operating systems require further protections beyond the limited device-file safeguards already implemented for Linux. They recommend extending capability checks to cover monitoring of a user's own files and any readable file on the system. For Windows specifically, the authors propose blocking the monitoring of entire drives, while both Windows and macOS should introduce kernel-level permission systems that incorporate context, access control, owned files and directories, and minifilters. The attacks work primarily in scenarios where a local attacker has an account with files readable by multiple users, though the list of globally readable files is extensive across platforms. Most attack scenarios depend on local access, but the demonstrated remote SSH keystroke capture shows the threats extend beyond physically shared computers to any system where users connect remotely. Organizations relying on file notification subsystems for legitimate application monitoring now face a choice between functionality and security until vendors implement comprehensive fixes that balance both needs. The research exposes a fundamental design oversight embedded in operating systems for up to two decades, revealing how features built for convenience can become exploitation vectors when permission models fail to anticipate side-channel risks.