adb
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONMETADATA_POISONING
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external, untrusted sources on an Android device, including system logs (
adb logcat), bug reports (adb bugreport), and file contents (adb shell cat,adb pull). A malicious application or file on the connected device could include instructions hidden in logs or text files intended to subvert the agent's behavior during a debugging session. - Ingestion points:
SKILL.md(instructions to useadb logcat,adb bugreport,adb shell cat, andadb shell dumpsysfor diagnostics) - Boundary markers: None identified; the skill does not instruct the agent to ignore instructions embedded in the device output.
- Capability inventory: The skill can execute various ADB commands to read system state, query content providers, and transfer files. It does not have network access or host-level file-write capabilities in the
allowed-toolsconfiguration. - Sanitization: No sanitization or filtering of content retrieved from the device is specified.
- [DATA_EXFILTRATION]: The skill provides tools for extracting sensitive information from a connected Android device. This includes the ability to query content providers (
adb shell content query), capture the device screen (adb shell screencap), read arbitrary files (adb shell cat,adb pull), and extract detailed system logs and bug reports. While standard for developer tools, these capabilities facilitate the exposure of sensitive user data. - [COMMAND_EXECUTION]: The skill utilizes shell command execution via the
adbutility. The execution is scoped to the subcommands explicitly listed in theallowed-toolsfrontmatter, which primarily focus on read-only and diagnostic operations. - [METADATA_POISONING]: There is a discrepancy between the capabilities documented in the markdown body and the permissions granted in the
allowed-toolsfrontmatter. The markdown lists 'Dangerous Commands' such asadb root,adb shell rm, andadb shell settings putas supported features requiring confirmation, but these commands are not included in theallowed-toolslist and would be blocked by the platform's security policy.
Audit Metadata