The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to read personal photos from a Google Drive synchronization folder located at ~/Library/CloudStorage/GoogleDrive-mike.cann@gmail.com/My Drive/Personal/Photos of Me/convex/Backgrounds Removed. These images are accessed via the view_image tool and the content is subsequently passed to an external image generation model as identity references. This process involves the exposure of personal data to an external AI service provider.
[COMMAND_EXECUTION]: The agent is instructed to execute the local bash script scripts/check-local-assets.sh. This script performs a discovery process on the local filesystem to identify absolute paths for Google Drive folders and specific image assets. Executing shell scripts provided within a skill allows for local system discovery and potential exploitation if the script were modified or if arguments were improperly handled.
[INDIRECT_PROMPT_INJECTION]: The workflow relies on processing untrusted data from external sources such as Notion pages, scripts, or video briefs to extract "hooks" for thumbnail concepts. This presents an attack surface where malicious instructions embedded in those documents could influence the agent's behavior.
Ingestion points: Notion pages, user scripts, and video briefs processed in the Core Workflow (SKILL.md).
Boundary markers: The prompt templates provided in SKILL.md and references/static-hosting-thumbnail.md do not utilize clear delimiters or instructions to ignore embedded commands in the external data.
Capability inventory: Local file reading, shell script execution (scripts/check-local-assets.sh), and access to image generation tools.
Sanitization: No sanitization or validation logic is applied to the extracted external content before it is interpolated into prompts for the image model.

mike-convex-thumbnail