The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines patterns to automatically trigger image generation, such as detecting placeholder.jpg or empty HTML sections. This process involves ingesting and analyzing untrusted data from the user's workspace. If an attacker controls parts of the source code being edited (e.g., through a malicious pull request or external dependency), they could embed instructions in comments or attributes that might influence the image generation prompt or the agent's behavior during the task. (Note: Category 8 severity is limited to LOW by standard analysis rules).
Ingestion points: Project files, specifically HTML/CSS/markdown files containing placeholders or empty visual elements.
Boundary markers: The skill does not explicitly define delimiters to separate inferred context from the system instructions.
Capability inventory: Executes a local CLI tool to send requests to external AI APIs.
Sanitization: No specific sanitization or validation of the inferred context is mentioned before it is used to craft the image generation prompt.
[COMMAND_EXECUTION]: The skill uses a local CLI script (node "${CLAUDE_PLUGIN_ROOT}/mcp-server/build/cli.bundle.js") to perform its core functionality. While this facilitates the intended image generation, it represents the execution of a binary/script within the agent's environment.

image-generation