The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes text from user-provided manuscripts and external PDFs to perform quote verification and claim grounding. Maliciously crafted content within these sources could potentially influence the agent's behavior during the verification cycle.
Ingestion points: The agent reads content from manuscript.docx, documents downloaded from Google Drive, and Readwise highlights.
Boundary markers: The instructions do not specify the use of delimiters or clear directives to ignore embedded instructions within the ingested text fragments.
Capability inventory: The skill possesses extensive capabilities, including network access (SSH, Google Drive API, Readwise, NLM), file system access, and the ability to execute shell commands and Python code.
Sanitization: There is no explicit process for sanitizing or escaping text extracted from external sources before it is passed to the LLM for analysis or interpolated into command templates.
[COMMAND_EXECUTION]: The skill executes multiple shell commands, such as rg, rga, and gws, to perform file searches and download operations. Notably, it uses ssh to connect to the WRDS service and execute SQL queries. The instructions for case verification suggest interpolating extracted plaintiff and defendant names directly into a shell command string, which creates a potential surface for command or SQL injection if the source manuscript contains adversarial strings.
[DYNAMIC_EXECUTION]: The skill utilizes uv run to execute a Python script for extracting footnotes from DOCX files. The script logic is provided as an inline string within the instructions. While functional for data extraction, the use of dynamic code execution at runtime increases the complexity and attack surface of the skill's environment.
[EXTERNAL_DOWNLOADS]: The skill fetches a bibliography file and PDF source documents from Google Drive, which is a well-known service. These operations target specific file IDs and utilize a provided email address (eddyhu@gmail.com) for the authentication context. These downloads are necessary for the skill's primary function of source verification.

source-verify