The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to analyze and score untrusted code provided by users, which introduces a surface for indirect prompt injection. This could allow instructions hidden within code snippets to influence the agent's scoring behavior or report output.
Ingestion points: The instructions in SKILL.md prompt the agent to process code directly from user input (e.g., 'Score this code on a 1-10 scale').
Boundary markers: The instructions lack explicit boundary markers or directives to treat input code as passive data, which increases the likelihood of the agent inadvertently following instructions embedded in the code.
Capability inventory: The skill is restricted to text processing and does not request access to shell execution, file system writes, or network operations, which significantly mitigates the impact of potential injections.
Sanitization: No sanitization or validation mechanisms are defined to filter out or escape potentially malicious instructions within the code being evaluated.

code-scoring