The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data that could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: According to SKILL.md, the agent accepts a "Source paper" (PDF or markdown) and a "Target database" (CSV or SAS files) as primary inputs.
Boundary markers: The instructions lack specific delimiters or "ignore embedded instructions" warnings for the agent when processing these external files.
Capability inventory: The skill uses the Bash, Write, and Edit tools, and is explicitly tasked with generating and running analysis scripts in Phase 3 and Phase 5.
Sanitization: There are no documented steps to sanitize or escape content extracted from source papers before it is interpolated into the code generation process.
[DYNAMIC_EXECUTION]: The skill generates and executes scripts at runtime based on the analysis of external documents.
Evidence: SKILL.md Phase 3 instructions include "Generate analysis code (Python with pandas + R via subprocess)" and Phase 5 requires verifying that "Code runs without errors on target data."
Context: While this is the primary purpose of the skill, generating code based on unstructured data interpreted by the model provides a vector for unintended code execution if the input source is adversarial.
[COMMAND_EXECUTION]: The skill leverages shell access and subprocess calls to perform its tasks.
Evidence: The Bash tool is included in the allowed-tools list in SKILL.md, and the workflow involves executing R scripts through Python's subprocess module.
Context: These tools are used for legitimate statistical analysis and data processing related to the skill's primary function.

replicate-study