The Agent Skills Directory

[PROMPT_INJECTION]: The subagent instructions for the grader, comparator, and analyzer agents are vulnerable to indirect prompt injection through the data they process.
Ingestion points: These agents ingest transcript.md, evals.json, and the contents of the outputs_dir as part of their core workflow.
Boundary markers: The instructions lack boundary markers or specific warnings to ignore instructions embedded within the ingested data.
Capability inventory: The agents have the ability to read and write files and execute shell commands (as specified in agents/grader.md).
Sanitization: There is no evidence of content sanitization or escaping before the data is processed by the subagents.
[COMMAND_EXECUTION]: The evaluation viewer component (eval-viewer/viewer.html) renders Excel data using innerHTML after conversion with the SheetJS library. This represents a potential Cross-Site Scripting (XSS) vector if a malicious spreadsheet is processed and viewed in the developer's local browser environment.
[COMMAND_EXECUTION]: The script eval-viewer/generate_review.py uses subprocess.run to execute the system lsof command to manage local network ports for the review server.
[EXTERNAL_DOWNLOADS]: The skill's HTML templates (assets/eval_review.html and eval-viewer/viewer.html) fetch JavaScript libraries and fonts from well-known services, including the SheetJS CDN and Google Fonts.

grok-skill-creator