The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides multiple bash scripts intended for execution in a developer's environment. These scripts utilize standard CLI tools including find, unzip, grep, xargs, dirname, sort, printf, and file to traverse the Gradle cache and inspect binary properties. While these are legitimate tools for the stated purpose of auditing, they grant the agent the ability to execute arbitrary commands in the local shell.- [INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface by ingesting and parsing external metadata from JAR archives in the local file system. If a dependency cache contains a maliciously crafted JAR file with specific pathnames, it could attempt to influence the agent's reasoning when it reviews the audit output.
Ingestion points: The audit loop in SKILL.md reads directory and filenames from ~/.gradle/caches/modules-2 using find and unzip -l.
Boundary markers: Absent. The audit results are printed to the console without delimiters or instructions to the agent to treat the output as potentially untrusted data.
Capability inventory: SKILL.md leverages shell execution (find, unzip, grep, file) to perform its audit functions.
Sanitization: The script uses grep -Eio '[^ ]+\.(so|dll|dylib)$', which acts as a filter by ensuring only specific file extensions are processed and displayed, limiting the injection surface to those patterns.

arm64-native-gap-audit