The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the execution of an external script ./extensions/banana/install.sh to install the necessary environment and MCP server. This script is not provided within the skill's source and its contents are unverifiable.
[COMMAND_EXECUTION]: The skill frequently instructs the agent to execute shell commands for image post-processing, metadata extraction, and cost tracking. This includes tools such as magick (ImageMagick), ffmpeg, cost_tracker.py, presets.py, and potrace.
[PROMPT_INJECTION]: The references/prompt-engineering.md file contains a section titled 'Safety Filter Rephrase Strategies' which provides explicit 'workarounds' and 'rephrase patterns' designed to bypass safety filters. These include techniques like abstraction, metaphor, and context shifting to generate content that might otherwise be blocked (e.g., violence, public figures).
[PRIVILEGE_ESCALATION]: Documentation in references/post-processing.md suggests the use of sudo apt install to install system dependencies, which involves elevating privileges.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it processes user-provided descriptions to construct complex prompts for the Gemini model.
Ingestion points: User-supplied descriptions for images in SKILL.md.
Boundary markers: The skill uses a 'Reasoning Brief' structure to delimit components, but lacks explicit sanitization of user input before interpolation into the generation pipeline.
Capability inventory: The agent can execute arbitrary shell commands (magick, ffmpeg) and interact with an external MCP server for image generation and editing.
Sanitization: There is no evidence of input filtering or escaping for user-controlled strings before they are used in shell commands or tool calls.

seo-image-gen