Task Manager can end, suspend and reprioritize processes, so it treats process control as privileged and treats the webview as untrusted input.

1. The backend enforces; the UI only explains. Every safety rule runs in Rust on every request. A compromised or buggy frontend cannot skip them.
2. No shell, no generic execution. There is no "run command" endpoint. Process control uses native APIs (TerminateProcess, SuspendThread, kill(2), setpriority(2)); no shell commands are constructed.
3. Minimal Tauri permissions. The capability file grants only core:default. Notifications, dialogs, autostart, window state and "open file location" are used from Rust only, so the frontend has no permission to call those plugins directly.
4. The app runs unprivileged. It never requests elevation. Operations the OS denies return a clear PermissionDenied explaining that administrator or root rights may be needed.
5. Local by default. No telemetry, and nothing is sent anywhere automatically. The only network access is the optional internet benchmark, which runs only when the user starts it, after a privacy notice and consent, and only to https://speed.cloudflare.com (see BENCHMARKS.md). Benchmark results stay on the machine.

Protection rules (core/src/process_control/protection.rs) are pure and unit-tested:

• Blocked: Task Manager itself; PID 0/4 (Windows kernel); PID 1 and kernel threads (Linux); processes Windows flags with IsProcessCritical; csrss, wininit, smss, services, lsass, winlogon, but only when their path is verified under %SystemRoot%\System32.
• Warning (requires explicit acknowledgement): critical names at unverified paths; svchost, explorer, dwm; root-owned core Linux daemons; root services started by init; raising priority to High.

Names alone never block. A process named lsass.exe in a Downloads folder gets a warning, not special protection.

Dangerous actions always need confirmation in the UI. Force end uses stronger wording, and the Delete key only opens the dialog.

• Command lines can contain passwords and tokens. They are:
  • never logged;
  • excluded from CSV/JSON exports unless the user ticks "Include command lines", which comes with a warning;
  • excluded from diagnostics unless the user opts in twice (process list, then command lines).
• Process environments are never read.
• Logs contain process names and PIDs for actions taken, never arguments.
• CSV exports neutralize spreadsheet formula injection: process names are attacker-controlled, so values starting with =, +, - or @ are prefixed with '.

default-src 'self'; style-src 'self' 'unsafe-inline'; img-src 'self' data:; connect-src ipc: http://ipc.localhost https://speed.cloudflare.com. No remote scripts, fonts or images are loaded. The single remote connect-src host is used only by the optional internet benchmark.

Benchmark commands accept enums, bounded numbers, GPU adapter ids and a mount point that must match the monitoring engine's list of writable volumes; there are no path, URL or command parameters. Storage tests write only .taskmanager-benchmark-<uuid>.tmp files (create-new) inside an app-owned folder that is canonicalized and must not be a symlink or escape the volume, keep a free-space reserve, respect a write limit, and delete their file on every exit path; the start-up sweep deletes only exact-pattern files in those folders. Internet-test results from the WebView are accepted only with a single-use token and are fully validated in Rust. Details: BENCHMARKS.md.

Please report vulnerabilities privately to the maintainers rather than in a public issue.