The Agent Skills Directory

[DATA_EXFILTRATION] (MEDIUM): Path traversal vulnerability in scripts/customize_template.py via the find_template function. The function joins assets_path with a user-provided template_name using the / operator. In Python's pathlib, if the second operand is an absolute path (e.g., /etc/passwd), it overrides the previous path, allowing the script to read arbitrary files from the system.
[DATA_EXFILTRATION] (MEDIUM): Arbitrary file write vulnerability in scripts/customize_template.py and scripts/validate_format.py. Both scripts accept user-provided output paths (--output and --report) without validation or sandboxing, enabling the agent to overwrite sensitive system files if prompted with a malicious path.
[COMMAND_EXECUTION] (LOW): Potential argument injection in scripts/validate_format.py. The script passes user-controlled file paths directly to subprocess.run as arguments for pdfinfo and pdffonts. While shell injection is avoided by using a list, an attacker could provide paths starting with hyphens to inject flags into the underlying system commands.
[PROMPT_INJECTION] (LOW): Indirect prompt injection surface in scripts/customize_template.py. The script interpolates unvalidated user input (title, authors, affiliations) directly into LaTeX templates. This could lead to malicious LaTeX code execution (e.g., using \write18 if the LaTeX environment is misconfigured) or influence downstream LLM processing of the generated documents.
Ingestion points: Command-line arguments (--title, --authors, etc.) and interactive prompts in scripts/customize_template.py.
Boundary markers: None present; input is directly substituted into the template string.
Capability inventory: File system write access via open().write().
Sanitization: None; the script uses raw re.sub for replacements without escaping special LaTeX or regex characters.

venue-templates