The Agent Skills Directory

[SAFE]: The skill's primary purpose is security observability testing. It explicitly reproduces common attack patterns—such as credential file access, base64 execution, and download-and-run—using benign marker payloads confined to a sandbox. This behavior is documented, intended, and directed by the user for testing purposes.
[COMMAND_EXECUTION]: The skill uses subprocess.run and shell scripts to exercise the agent's command execution capabilities. These operations are essential for generating the telemetry it aims to test and are restricted to a temporary run directory owned by the skill.
[EXTERNAL_DOWNLOADS]: The skill performs network probes using curl and uvx. These target well-known public endpoints (example.com) or a bundled loopback HTTP fixture (assets/http-server.py) for local verification. No sensitive data is exfiltrated to untrusted external domains.
[OBFUSCATION]: The skill includes a 'decode-and-execute' probe (probe-security.sh) that uses a base64-encoded string. This is a deliberate test case to verify if telemetry systems can detect encoded command execution; the decoded payload is a harmless printf command.

harness-check