The Agent Skills Directory

[PROMPT_INJECTION]: The ai-security skill components (such as ai_threat_scanner.py and atlas-coverage.md) contain numerous strings representing prompt injection and jailbreak attacks. These include instructions to "ignore all previous instructions," "reveal system prompt," and act as "DAN" (Do Anything Now). These strings are used as seed prompts for security scanning and examples of MITRE ATLAS techniques, which is consistent with the skill's primary purpose of AI security assessment.
[COMMAND_EXECUTION]: Several scripts in the collection use the Python subprocess module to execute system commands. For instance, pr_analyzer.py executes git diff to identify code changes, and gws_doctor.py executes the gws CLI tool to verify Google Workspace connectivity. This is standard functional behavior for the suite's development and administrative tools.
[DYNAMIC_EXECUTION]: The review_report_generator.py script programmatically executes other local Python scripts (pr_analyzer.py and code_quality_checker.py) using sys.executable. The execution paths are constructed from the script's own directory, which is a standard pattern for tool aggregation in this suite.
[INDIRECT_PROMPT_INJECTION]: Several skills, particularly code-reviewer and ai-security, are designed to process untrusted external data (such as git diffs or external RAG content). The ai-security documentation explicitly identifies this as a high-risk vector and provides guidance on treating retrieved content as untrusted input. The presence of this attack surface is expected for a security auditing tool and is accompanied by appropriate warnings.

engineering-skills