The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the adb (Android Debug Bridge) command-line tool to interact with connected devices. It includes shell pipelines and command substitutions (e.g., adb logcat --pid=$(adb shell pidof pkg)), which could be misused if arguments are not properly sanitized during agent interaction.
[INDIRECT_PROMPT_INJECTION]: The skill provides tools to ingest and process external data that the AI agent does not control.
Ingestion points: adb logcat, adb logcat -b crash, and adb pull /data/anr/traces.txt read data directly from a connected Android device.
Boundary markers: There are no explicit instructions or delimiters provided to the agent to ignore or sanitize instructions that might be embedded within the log output of a malicious or compromised application.
Capability inventory: The skill allows for significant device interaction, including adb install, adb shell pm clear, and adb shell am start, creating a risk if the agent is manipulated by injected content in logs.
Sanitization: No sanitization or filtering logic is present for the logs being read.
[DATA_EXFILTRATION]: The skill facilitates the extraction of sensitive information from connected peripherals.
Evidence: Commands like adb exec-out screencap -p > screen.png capture the device's screen, and adb pull /data/anr/traces.txt retrieves system-level application traces. While these are primary functions of a debugging skill, they expose user data from the device to the agent's context.
[EXTERNAL_DOWNLOADS]: The install_command in the metadata fetches the skill definition from a public repository on GitHub.
Evidence: curl -sL "https://raw.githubusercontent.com/majiayu000/claude-skill-registry-data/main/data/android-adb/SKILL.md" > .claude/skills/android-adb/SKILL.md. This is documented as a standard installation procedure for this registry.

android-adb