pptx
Warn
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: MEDIUMPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The scripts
ooxml/scripts/unpack.pyandooxml/scripts/validation/redlining.pyare vulnerable to Directory Traversal (Zip Slip). These components utilizezipfile.ZipFile.extractall()to extract PowerPoint content (which are ZIP archives) without validating that the internal file paths remain within the target destination. An attacker could craft a malicious presentation containing members with traversal sequences (e.g.,../../.ssh/authorized_keys) to overwrite arbitrary files and potentially gain unauthorized access or persistence. - [COMMAND_EXECUTION]: Multiple components execute system-level utilities through
subprocess.run(). Specifically,ooxml/scripts/pack.pyandscripts/thumbnail.pycallsoffice(LibreOffice) for file conversion, whilescripts/thumbnail.pyusespdftoppmfor image generation andooxml/scripts/validation/redlining.pyusesgitfor diffing. While implemented using argument lists rather than raw strings to reduce injection risks, these operations execute with the agent's privileges on data derived from potentially untrusted inputs. - [DYNAMIC_EXECUTION]: The
scripts/html2pptx.jscomponent uses the Playwright browser engine to render HTML content for slide generation. This initiates achromium.launch()call to process HTML/CSS layouts. If an agent is induced to process untrusted HTML from a prompt or external file, this mechanism could be exploited to execute scripts or access local resources within the headless browser environment. - [INDIRECT_PROMPT_INJECTION]: The skill has a high vulnerability surface for indirect prompt injection due to its extensive processing of external data from PowerPoint and HTML files. Ingestion points include text extraction via
markitdownandscripts/inventory.py, and HTML rendering via Playwright. The absence of explicit boundary markers or sanitization logic increases the risk of the agent obeying malicious instructions embedded in slide content or HTML structures, leveraging the skill's significant local capabilities.
Audit Metadata