The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted user data (such as 'pathway_description', 'workflow_description', and 'research_topic') and interpolates it into predefined scenario templates without sanitization or boundary markers. This could allow an attacker to influence the final prompt generated by the agent, potentially leading to the consumption of image generation credits for unintended purposes.
Ingestion points: The skill uses 24 different scenario templates (e.g., references/templates/01-signaling-pathway.json) that include placeholders for user-supplied descriptions.
Capability inventory: The skill can execute local Python scripts (scripts/generate_image.py and scripts/generate_geometry.py) which perform network requests and write files to disk.
Boundary markers: No explicit boundary markers (e.g., delimiters) or "ignore embedded instructions" warnings are used when interpolating user data into prompts.
Sanitization: No escaping or validation of external content is performed before interpolation into the final prompt.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to external image generation APIs (like OpenAI or apimart.ai) as configured by the user. The scripts/generate_image.py script sends the user-provided IMG_API_KEY in the Authorization header to the IMG_BASE_URL defined in the environment. This is the intended functionality but relies on the security of the user-provided endpoint.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute local Python scripts to fulfill its primary functions. This includes scripts/generate_image.py for API-based image generation and scripts/generate_geometry.py for creating mathematically accurate plots using the matplotlib library.

stem-illustration