The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell commands including cat, jq, find, and ls to locate and read benchmark files. The file paths are dynamically constructed using the $ARGUMENTS variable (e.g., <plugin>/skills/<skill>/eval-results/benchmark.json). If the user-provided target argument contains path traversal sequences like ../, the skill may attempt to access files outside the intended evaluation directories.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8) because it ingests and processes untrusted data from external JSON files (benchmark.json, history.json, plugin-benchmark.json).
Ingestion points: Files are read from the filesystem using cat and jq in Step 2 and the 'Agentic Optimizations' section.
Boundary markers: The skill lacks explicit delimiters or warnings to the agent to ignore instructions embedded within the JSON content.
Capability inventory: The skill has access to shell execution (Bash), file system navigation (find, ls, Glob), and file reading (Read, cat, Grep).
Sanitization: There is no evidence of validation or sanitization of the JSON content before it is displayed to the agent, allowing malicious metrics or descriptions to potentially influence downstream logic.

evaluate-report