The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources which could contain malicious instructions.
Ingestion points: The skill reads data from .accdb, .mdb, and .csv files using pyodbc (Python), mdbtools (Bash), and DoCmd.TransferText (VBA).
Boundary markers: No explicit boundary markers or "ignore instructions" wrappers are suggested for the ingested content.
Capability inventory: The skill utilizes powerful capabilities including executing SQL via CurrentDb.Execute, running VBA macros, and invoking shell commands through msaccess.exe and mdbtools.
Sanitization: The skill provides strong sanitization guidance, explicitly instructing the use of parameterized DAO.QueryDef objects and escaping apostrophes in dynamic filters to prevent SQL injection.
[COMMAND_EXECUTION]: The skill instructions involve executing external tools and binaries.
Evidence: Provides examples for driving msaccess.exe with command-line switches (/x, /compact, /cmd) and using mdbtools utilities (mdb-tables, mdb-export, mdb-schema) for database management on Linux and macOS.
[PRIVILEGE_ESCALATION]: The skill contains instructions for installing software that requires elevated permissions.
Evidence: Example 3 in SKILL.md suggests using sudo apt install mdbtools to set up the environment on Linux systems.

ms-access