The Agent Skills Directory

[COMMAND_EXECUTION]: A path traversal vulnerability (Zip Slip) exists in ooxml/scripts/unpack.py, ooxml/scripts/validation/redlining.py, and ooxml/scripts/validation/base.py. These scripts use zipfile.extractall() on PowerPoint files without validating that archive members remain within the target directory. A malicious PowerPoint archive could write or overwrite files outside the extraction path (e.g., via ../../.bashrc).\n- [COMMAND_EXECUTION]: The skill frequently invokes external system utilities using subprocess.run, including soffice, pdftoppm, and git. These operations process user-provided files and paths, representing a risk if input filenames are not properly sanitized or if the underlying binaries are exploited.\n- [REMOTE_CODE_EXECUTION]: The scripts/html2pptx.js utility uses Playwright to launch a headless Chromium browser instance to render slides. This provides a complex execution environment where untrusted HTML or CSS could potentially be used to trigger browser-based exploits, access local files, or perform network requests if not strictly sandboxed.\n- [PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection. It extracts all text content from presentations and instructs the agent to read these outputs in their entirety. Malicious instructions embedded within a slide could potentially hijack the agent's logic during analysis or summarization tasks.

pptx