The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The file gpt.md contains a hardcoded password string (SecureRFCPassword123!) within a configuration dictionary (SAP_CONFIG) intended for SAP NetWeaver RFC connections.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection attacks.
Ingestion points: The sap_gui_scraper.py script in SKILL.md and related logic in gemini.md ingest data from SAP ALV Grid controls (GuiGridView) and status bars (sbar). These controls often contain user-generated content from the ERP system.
Boundary markers: The instructions do not provide boundary markers or explicit guidance for the agent to treat data read from the SAP system as untrusted.
Capability inventory: The skill possesses significant capabilities, including the ability to manipulate the SAP GUI desktop application via the COM API (win32com.client), navigate transactions, and modify fields.
Sanitization: There is no evidence of sanitization or validation of the data retrieved from SAP grids before it is printed or potentially processed in subsequent agent steps.
[COMMAND_EXECUTION]: The skill provides scripts that automate desktop application behavior via the SAP GUI Scripting API. Notably, the Operational Directive in SKILL.md instructs the user to uncheck 'Notify when a script attaches to SAP GUI'. This action disables a native security warning, allowing the agent to perform actions in a headless, non-transparent manner which increases the risk of undetected malicious activity.
[DATA_EXFILTRATION]: The Python script in gpt.md demonstrates network connectivity to an external host (sap-app-server.enterprise.io) via the pyrfc library. While necessary for the skill's function, it establishes a network exit point for data read from the local environment or the SAP system.

sap-gui