The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill incorporates a specific attack surface by ingesting external data from web searches and user-supplied blog drafts. However, it implements effective mitigations through detailed instructions (e.g., in rules/repair-method.md and adversarial test cases in assets/evals/naver-blog-maker-cases.jsonl) that mandate treating external inputs as data rather than executable instructions. Sanitization rules for HTML output further reduce this risk.
[DATA_EXFILTRATION]: No patterns of data exfiltration were detected. The skill operates entirely without user credentials or privileged access. It uses standard search and page-fetching tools for research and delivers all results directly to the user. Instructions strictly forbid any automated engagement actions on the Naver platform.
[REMOTE_CODE_EXECUTION]: The skill is completely self-contained and does not download or execute any external scripts or packages. It relies solely on internal Markdown rules and references.
[DYNAMIC_EXECUTION]: While the skill generates HTML for 'rich' post formatting, it includes explicit security constraints to exclude scripts, event handlers, and Base64 images. The static HTML template provided for the copy-to-clipboard functionality uses standard, safe DOM APIs.
[PROMPT_INJECTION]: The skill contains built-in defenses against prompt injection, including adversarial evaluation cases and instructions that prevent the AI from bypassing its own safety guidelines, such as rules against inventing experiences or omitting sponsorship disclosures.

naver-blog-maker