The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted SVG files which can contain malicious instructions within text labels or metadata intended to influence the agent's review findings or repair steps. * Ingestion points: The SVG file is read and parsed by scripts/svg_check.py. * Boundary markers: There are no explicit instructions for the agent to ignore natural language content found within the SVG during its analysis phase. * Capability inventory: The agent has capabilities to execute the Python script, run rendering commands in the shell, and use a browser. * Sanitization: The tool flags technical risks like scripts but does not provide specific sanitization against prompt injection in visual labels.
[COMMAND_EXECUTION]: The skill instructs the agent to probe the local environment for SVG rendering tools and execute them via shell commands (e.g., rsvg-convert, inkscape, magick). This relies on executing binary tools based on environment discovery.
[DYNAMIC_EXECUTION]: The scripts/svg_check.py script uses xml.etree.ElementTree to parse SVG data, which is not secure against malicious XML constructs. The script attempts to block XXE attacks by searching for entity declarations in the first 4096 bytes of the file, but this check can be bypassed by an attacker using padding or large comments, potentially allowing for local file access or unexpected behavior.

check-svg-diagrams