The Agent Skills Directory

[Indirect Prompt Injection] (MEDIUM): The script extracts and prints metadata (page size, fonts, image details) from external PDF files without sanitization. A malicious PDF could contain instructions in its metadata (e.g., font names or document info) that are then echoed into the agent's context, potentially leading to agent subversion.
Ingestion points: File content/metadata processed via pdfinfo, pdffonts, and pdfimages in review_poster.sh.
Boundary markers: None. Extracted data is printed directly to the output.
Capability inventory: The script performs read-only operations and subprocess calls to local utilities; it does not have network or write capabilities itself, but its output influences the agent.
Sanitization: None. Metadata is printed as raw strings.
[Command Execution] (LOW): The script executes local system binaries (pdfinfo, pdffonts, pdfimages, ls, gs). While these are standard tools, they must be pre-installed on the host system (e.g., via poppler-utils).
[Terminal Injection] (LOW): The script prints raw output from PDF tools. If a PDF contains ANSI escape sequences in its metadata, it could manipulate the terminal display of a human user.

latex-posters