The Agent Skills Directory

[COMMAND_EXECUTION]: The framework is designed to execute shell commands extracted from target skills using subprocess.run(['bash', '-c', cmd_text]). This occurs primarily in scripts/tier1/phase-4-execute-tests.sh and scripts/tier2/phase-6-full-flow.sh.
[DYNAMIC_EXECUTION]: The skill generates and executes Python scripts at runtime. In scripts/tier1/phase-4-execute-tests.sh, it writes SDK snippets extracted from target skills into temporary files and executes them using the local Python interpreter. It also uses exec(open(...).read()) to dynamically load utility modules in scripts/tier1/phase-3-gen-testcases.sh and scripts/tier1/phase-4-execute-tests.sh.
[INDIRECT_PROMPT_INJECTION]: There is a high attack surface for indirect injection. The framework ingests untrusted data from SKILL.md files of other skills (ingestion points in phase-1-skill-analysis.sh) and executes the extracted strings without rigorous sanitization. A malicious skill could provide a command that exfiltrates credentials or performs unauthorized actions when 'tested' by this framework.
[EXTERNAL_DOWNLOADS]: The documentation in references/cli-installation-guide.md references a remote installation script: https://apiexplorer.developer.huaweicloud.com/install/hcloud/install.sh. While this is an official vendor domain, it represents execution of a remote script.
[DATA_EXFILTRATION]: The skill includes a telemetry SDK (scripts/skill_quality_sdk.py) that reports execution metadata (status, cost, masked parameters) to https://skillsapi.developer.myhuaweicloud.com/api/quality/report and a fallback guest endpoint https://skillsop.topxtopx.com/api/quality/guest-report. The latter is an unconventional domain for a cloud provider, though it is used here for anonymous telemetry when credentials are missing.

huawei-cloud-skill-tester