The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/adb_helper.sh passes user-provided arguments (such as package names, component names, and IP addresses) into shell commands via adb shell and adb connect. While variables are quoted, they are not sanitized against subshell execution or special characters, creating a risk of command injection on either the host machine or the connected Android device.
[DATA_EXFILTRATION]: The skill provides functions (cmd_dump_prefs and cmd_dump_db) to extract private application data, including SharedPreferences and SQLite databases, from the connected Android device. This allows the agent to access potentially sensitive user information stored within third-party apps.
[COMMAND_EXECUTION]: The script attempts to execute commands with root privileges (su -c) on the connected device to access restricted files and directories.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data.
Ingestion points: The logcat functions in scripts/adb_helper.sh read and display real-time logs from the connected Android device.
Boundary markers: None. Logs are displayed directly in the agent context without delimiters or warnings.
Capability inventory: The skill has extensive capabilities including file system access (via adb pull/rm), application management (install/uninstall), and arbitrary command execution on the device (adb shell).
Sanitization: No filtering or sanitization is performed on the log output. An attacker could trigger logs containing malicious instructions to influence the agent's behavior.

android-adb-toolkit