The Agent Skills Directory

[DATA_EXFILTRATION]: The ConstructionEmailSender class in SKILL.md provides a mechanism to read files from the local filesystem using open(attachment_path, 'rb') and transmit them as email attachments. A malicious actor could exploit this by providing paths to sensitive system files (e.g., SSH keys, configuration files), leading to data exfiltration via the configured SMTP server.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data from Excel (RFI_Log.xlsx), CSV, and JSON formats as described in instructions.md.
Ingestion points: Data is loaded through pandas.read_excel and other format-specific parsers.
Boundary markers: There are no explicit boundary markers or instructions to ignore embedded commands within the ingested data.
Capability inventory: The skill possesses the ability to read files from the filesystem and send data over the network via SMTP.
Sanitization: The provided Python snippets do not include sanitization or validation logic for the content or file paths provided in the input data, creating a surface for injection attacks.
[COMMAND_EXECUTION]: The skill relies on the execution of Python code to generate templates and handle email logic. While the logic itself is not inherently malicious, the lack of input validation on user-supplied parameters that flow into file operations and network requests increases the risk of the agent performing unintended actions.

email-construction