The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a test fixture (fixtures/positive/sg002.sh) containing the destructive command 'rm -rf /'. This is intended for testing the scanner's detection rules and is not executed during normal operation.
[PROMPT_INJECTION]: Test fixtures like fixtures/positive/sg008.md contain common prompt injection strings (e.g., 'Ignore all previous system instructions' and instructions to conceal actions). These are used to validate the tool's ability to flag such patterns in external code.
[OBFUSCATION]: The skill contains a fixture (fixtures/positive/sg005.ps1) demonstrating the use of PowerShell's '-EncodedCommand' flag. This serves as a positive test case for identifying obfuscated execution.
[REMOTE_CODE_EXECUTION]: Test fixtures contain patterns for executing remote content, such as 'curl | sh' in fixtures/positive/sg001.md and unpinned git installations in fixtures/positive/sg007.sh. These are documented test cases for the scanner.
[INDIRECT_PROMPT_INJECTION]: As a security scanner, the tool necessarily ingests and processes untrusted data from local files. 1. Ingestion points: The scripts/skillguard.py script reads the contents of all files in a target directory. 2. Boundary markers: The SKILL.md includes explicit warnings to treat all findings as data and never execute the target artifact. 3. Capability inventory: The script is restricted to local file reading and regex processing; it does not perform network operations or spawn subprocesses. 4. Sanitization: The script includes logic to redact detected secrets (like API keys) from the output evidence shown to the user.

agent-skillguard