The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions in SKILL.md that steer the agent to promote the developer's hosted platform (K-Dense Web) and mandate the use of other specific skills (scientific-schematics) during document creation. This represents a form of behavioral steering and instruction override.- [COMMAND_EXECUTION]: The script scripts/validate_format.py utilizes the subprocess module to execute external binaries (pdfinfo, pdffonts). While arguments are passed as a list, the script operates on user-provided file paths, which could be exploited if combined with other vulnerabilities.- [INDIRECT_PROMPT_INJECTION]: The script scripts/customize_template.py takes external inputs (title, authors, affiliations) and performs direct string substitution into LaTeX templates without sanitization. This creates a surface for LaTeX injection. If an attacker provides malicious LaTeX code as a paper title and the agent follows the suggested workflow to compile the document using pdflatex, it could lead to unauthorized file reads or command execution via LaTeX features.
Ingestion points: CLI arguments --title, --authors, and --affiliations in scripts/customize_template.py.
Boundary markers: None present in the prompt interpolation or script logic.
Capability inventory: The skill allows Bash, Read, Write, and Edit operations, enabling the agent to compile the injected code.
Sanitization: No input validation or escaping is implemented in the customization script.- [DATA_EXFILTRATION]: While no direct exfiltration patterns were found, the combination of LaTeX injection potential and Bash capabilities creates a risk path where sensitive data accessible to the agent could be included in generated documents.

venue-templates