The Agent Skills Directory

[METADATA_POISONING]: The skill instructions in SKILL.md include a link (https://chatgpt-plus.top/list/#/home) that has been identified as a phishing URL by automated security scans. This URL is explicitly recommended for users in mainland China.
[COMMAND_EXECUTION]: The skill instructs the agent to execute bundled Python scripts (scripts/generate.py and scripts/online_link.py) to handle image generation and link creation, which involves running shell commands within the agent's environment.
[DATA_EXFILTRATION]: The scripts/generate.py script transmits user-provided image files and generated prompts to external API endpoints. While it defaults to api.openai.com, the skill also suggests tryallapi.com and allows the endpoint to be overridden via the GPTIMAGE_BASE_URL environment variable, potentially directing sensitive data to untrusted servers.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill processes untrusted user-supplied product reference photos and text descriptions as defined in SKILL.md.
Boundary markers: The skill uses a 'PRODUCT IDENTITY LOCK' to structure product data, but it does not provide explicit instructions to the agent to ignore or sanitize commands that might be embedded within the user-provided images or text.
Capability inventory: The skill has the capability to execute scripts, perform network operations (API calls), and write files to the local disk.
Sanitization: There is no evidence of input validation or sanitization for instructions that could be hidden in the processed data.
[DYNAMIC_EXECUTION]: The script scripts/generate.py performs dynamic assembly of multipart form data and decodes base64-encoded image data returned from remote APIs at runtime.

lifestyle-scene