academic-defense-beamer
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted LaTeX source files from user-provided thesis repositories, which creates a surface where embedded instructions could attempt to influence agent behavior during the planning or Q&A preparation phases.
- Ingestion points:
scripts/extract_thesis.pyandscripts/tex_loader.pyread content from the user's thesis repository. - Boundary markers: Present.
SKILL.mdandagents/qa-committee-agent.mdcontain explicit instructions for the agent to treat thesis data as untrusted and to ignore any commands found within it. - Capability inventory: The skill uses
subprocess.runto execute local commands inscripts/build_deck.py,scripts/typst_backend.py, andscripts/check_typst.py. - Sanitization: Present. The skill implements dedicated escaping logic (
tex_escape,ty_literal) inscripts/build_deck.pyandscripts/typst_backend.pyto sanitize untrusted content before inserting it into output templates. - [COMMAND_EXECUTION]: The skill executes local system commands to compile LaTeX and Typst documents and to query document metadata.
- Evidence:
scripts/build_deck.py,scripts/typst_backend.py, andscripts/check_typst.pyuse thesubprocess.runfunction to invokelatexmkandtypstbinaries. - Context: These executions are fundamental to the skill's primary purpose of generating PDFs. The skill constructs these commands using controlled logic and includes checks to ensure only local, pre-installed packages are used.
Audit Metadata