The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content, including article text, Markdown files, and video subtitles (.srt, .vtt), to generate core summaries and image prompts. \n- Ingestion points: Workflow defined in SKILL.md accepts raw text or local file paths from the user. \n- Boundary markers: Absent. The skill lacks instructions for the agent to use delimiters or warnings to ignore commands embedded within the input documents. \n- Capability inventory: Invokes image generation tools (e.g., DALL-E 3, FPBrowser2API) and performs text summarization. \n- Sanitization: Absent. There is no logic to escape or validate the contents of the provided files before processing. \n- [DATA_EXFILTRATION]: The skill workflow explicitly encourages the agent to accept local file paths (e.g., D:\path\to\article.md) for content analysis. Without file-type or content validation, an attacker could provide paths to sensitive files (such as .env, .ssh/id_rsa, or credential configs) which the agent might then summarize and leak in its output as part of the 'cover design' summary. \n- [COMMAND_EXECUTION]: The 'Post-Processing SOP' section mentions the use of 'scripts' and 'automation tools' for image cropping and delivery. While no scripts are provided in the skill itself, the reliance on external execution environments for core functionality introduces a risk of uncontrolled command execution if the user environment is not properly secured.

article-cover-designer