The Agent Skills Directory

[COMMAND_EXECUTION]: Several scripts, including google-workspace-cli/scripts/gws_recipe_runner.py, google-workspace-cli/scripts/auth_setup_guide.py, and code-reviewer/scripts/review_report_generator.py, utilize subprocess.run with shell=True or execute other Python scripts via system calls. This approach is highly vulnerable to command injection if placeholders or arguments are populated with unsanitized data from the agent's context or ingested files.
[INDIRECT_PROMPT_INJECTION]: The collection features a wide variety of tools designed to process untrusted data from the local environment and external sources, posing a high risk for indirect prompt injection attacks:
Ingestion points: a11y-audit/scripts/a11y_scanner.py (scans codebases), code-reviewer/scripts/pr_analyzer.py (analyzes git diffs and commit messages), ai-security/scripts/ai_threat_scanner.py (reads prompt test data from external files), and incident-commander/scripts/timeline_reconstructor.py (processes event logs).
Boundary markers: The scripts lack explicit delimiters or safety instructions (e.g., 'ignore embedded instructions') when interpolating untrusted content into the model's context.
Capability inventory: The skills possess significant execution capabilities, including shell access (google-workspace-cli), file system writes, and configuration management for cloud environments.
Sanitization: No evidence of robust input validation, escaping, or filtering was found in the data processing pipelines.
[PROMPT_INJECTION]: Skill components in the ai-security module (SKILL.md, references/atlas-coverage.md) contain explicit examples of prompt injection techniques, such as 'DAN' (Do Anything Now) mode framing and system prompt extraction. While these are intended for security assessment purposes, they constitute high-risk instructional patterns.
[METADATA_POISONING]: There is a notable discrepancy between the authoritative author name provided in the analysis context ('thewiningturtle') and the names found in the skill metadata and documentation ('Alireza Rezvani' and 'Abbas Mir'). This inconsistency could be indicative of deceptive metadata designed to misrepresent the provenance or intended behavior of the skill collection.

engineering-skills