The Agent Skills Directory

[DATA_EXFILTRATION]: The scripts/generate.py script contains an upload_local_paths function that automatically uploads local files to fal.ai's cloud storage if their paths are included in manifest fields ending in _url or _urls. While this is a functional requirement for processing image assets, it creates a mechanism where the agent could inadvertently or maliciously exfiltrate sensitive files (e.g., SSH keys or configuration files) if manipulated to include those paths in the generation manifest.
[COMMAND_EXECUTION]: The skill instructions (SKILL.md) direct the agent to execute local Python scripts (scripts/generate.py and scripts/gallery.py) to manage API interactions and generate a local HTML gallery. This involves running code within the user's local environment using a dedicated virtual environment.
[PROMPT_INJECTION]: The skill processes user-supplied brand documentation files (brand/brand-dna.md, brand/brand-voice.md, brand/icp-cards.md) to guide the generation of ad copy. This creates a surface for indirect prompt injection attacks.
Ingestion points: The agent reads markdown files from the brand/ directory to contextualize its generations.
Boundary markers: Absent. The instructions command the agent to read all files but do not provide specific delimiters or instructions to disregard potential malicious directives embedded within those files.
Capability inventory: The skill has the capability to read and write files, perform network operations via the fal-client library, and execute local Python scripts.
Sanitization: Absent. Content from the brand files is directly used by the LLM to influence ad copy variations and image briefings.

ad-creative-engine