The markdown_table_to_image tool accepts a caller-controlled path parameter and passes it directly to get_html_table_image:
# pptagent/mcp_server.py:127-143
def markdown_table_to_image(markdown_table: str, path: str, css: str) -> str:
"""
Args:
path (str): The file path where the image will be saved
"""
html = markdown_to_html(markdown_table)
get_html_table_image(html, path, css) # ← no path validation
return f"Markdown table converted to image and saved to {path}"
# pptagent/utils.py:337-366
def get_html_table_image(html: str, output_path: str, css: str = None):
parent_dir, base_name = os.path.split(output_path)
if parent_dir and not os.path.exists(parent_dir):
os.makedirs(parent_dir) # ← creates arbitrary directories
hti = Html2Image(...)
hti.screenshot(
html_str=html,
css_str=css,
save_as=base_name, # ← writes image to any directory
size=(1000, 600),
)
os.makedirs(parent_dir) creates arbitrary directory trees, and Html2Image.screenshot writes the rendered image to parent_dir/base_name. Unlike download_file in the same project, there is no is_relative_to(workspace) guard. This behaviour can be fixed with the same pattern as the above.
The concrete attack scenarios include
path = "/home/user/.ssh/authorized_keys" — replaces the authorized_keys file with an image binary (breaks - SSH but could be an image crafted with a specific PNG/JPEG payload).path = "/var/www/html/uploads/shell.php" — writes the rendered PNG there; the file has the .php extension but PNG content; combined with Apache Options +MultiViews or file-include vulnerabilities could be dangerous.path = "/root/test/probe.png" — if the directory is created, confirms the target path exists; if it errors, reveals permissions information.