The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources which could potentially contain malicious instructions intended to influence the agent's behavior.
Ingestion points: As described in the Workflow section of SKILL.md, the skill uses the web_fetch tool to retrieve content from URLs and reads user-provided files such as PDFs, images, and documents.
Boundary markers: The instructions lack specific delimiters or constraints that would isolate retrieved external content from the model's instructions, and they do not explicitly warn the model to ignore instructions embedded within the processed data.
Capability inventory: The agent has capabilities to perform network fetches via web_fetch, read local files, and write configuration files to the output directory using the present_files tool.
Sanitization: No specific sanitization, filtering, or validation logic is defined for the retrieved web content or document text before it is processed for brand extraction.

brand-yml