The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources such as vulnerability findings (YAML) and Pentest Rules of Engagement (Markdown). To mitigate risks of malicious content influencing the generated diagrams or breaking the Markdown structure, the skill implements a strict mlabel sanitization function. This function strips newlines, quotes, backticks, angle brackets, and various bracket types from labels before they are included in the Mermaid output.
[DATA_EXFILTRATION]: The skill reads security configuration and permission manifests from the user's home directory (~/.claude) to generate an 'AI Agent Risk Graph'. While this involves sensitive configuration data, the skill has no network access (explicitly forbidden via forbidden-tools: WebFetch) and only writes the resulting visualizations to the local project directory, posing no risk of exfiltration.
[COMMAND_EXECUTION]: The skill uses a local Node.js script to process data and generate diagrams. The script is written in pure Node.js with no external dependencies and uses standard filesystem operations for reading artifacts and writing reports.
[PRIVILEGE_ESCALATION]: Access to the ~/.claude directory is restricted to reading the agent's own configuration and manifests, which is consistent with the skill's purpose of visualizing the agent's current security posture.

security-viz