The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses the | safe filter in several Jinja2 templates (such as research_report.html and backtest_report.html) to render HTML content for tables and position changes, bypassing default auto-escaping. If the data source for these tables is influenced by an attacker, it could lead to XSS.\n
Ingestion points: Data for reports is ingested via the ResearchReport dataclass in research_report.py and the context passed to the render method in renderer.py.\n
Boundary markers: No specific boundary markers or instructions were found in SKILL.md or the templates to delimit untrusted data during interpolation.\n
Capability inventory: The skill possesses file-write capabilities in renderer.py (save and to_pdf methods) and file-read capabilities in research_report.py (_read_figure_png method).\n
Sanitization: While Jinja2 auto-escaping is enabled in renderer.py, the | safe filter in templates like research_report.html explicitly bypasses this sanitization for selected fields.

report