The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides instructions for executing commands with administrative privileges using sudo. This includes updating system packages and copying files to protected system directories. Evidence: The skill instructs the agent to run 'sudo tlmgr update --self', 'sudo tlmgr install tcolorbox tikz pgf', 'sudo cp assets/medical_treatment_plan.sty /usr/local/texlive/texmf-local/tex/latex/', and 'sudo texhash'.
[DYNAMIC_EXECUTION]: The skill passes natural language user input directly into a shell command context, creating a surface for command injection. Evidence: The instruction 'python scripts/generate_schematic.py "your diagram description" -o figures/output.png' allows for arbitrary shell command execution if the input string is not properly sanitized.
[COMMAND_EXECUTION]: The skill relies on the Bash tool to execute local Python scripts and LaTeX compilation tools. Evidence: Extensive use of 'pdflatex', 'xelatex', 'bibtex', and various local scripts like 'check_completeness.py', 'validate_treatment_plan.py', and 'timeline_generator.py'.
[INDIRECT_PROMPT_INJECTION]: The skill ingests clinical data and patient descriptions to generate documents, providing a vector for indirect prompt injection. 1. Ingestion points: Patient demographics, clinical diagnoses, and diagram descriptions provided via natural language. 2. Boundary markers: The instructions lack delimiters or explicit directives to ignore instructions embedded within the medical data. 3. Capability inventory: The skill is allowed access to Bash, Read, Write, and Edit tools, which could be abused if the agent follows malicious instructions in the input data. 4. Sanitization: There is no evidence of sanitization or validation of external content before its use in scripts or document templates.

treatment-plans