The Agent Skills Directory

Indirect Prompt Injection: The skill is designed to read and interpret content from user-provided CSV files. This creates a potential surface for indirect prompt injection if the source data contains malicious instructions intended to influence the agent's behavior during processing.
Ingestion points: Untrusted data enters the agent context through the file specified in the csv-path parameter as described in SKILL.md.
Boundary markers: The instructions do not explicitly require the use of boundary markers or delimiters to isolate data from the CSV, which could lead the agent to inadvertently follow instructions embedded in the file content.
Capability inventory: The skill utilizes local helper scripts (report_tools.sh and report_tools.ps1) to perform lookups and validation, and it has the capability to write JSON reports to the workspace.
Sanitization: The skill employs helper scripts to perform structural and consistency validation of the output report against a predefined JSON schema.
Command Execution: The skill includes helper scripts and a test suite that execute shell commands and subprocesses to perform its core tasks.
Evidence: The Python test script tests/test_report_tools.py uses subprocess.run() to invoke the Bash and PowerShell scripts during automated testing.
Evidence: The helper scripts (scripts/report_tools.sh and scripts/report_tools.ps1) execute commands such as jq and ConvertFrom-Json to handle JSON data and validation. These executions are confined to local scripts provided within the skill.

assessment-report-converter