The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The 'URL Reader' example in SKILL.md demonstrates a functional pattern that ingests untrusted data from the internet and interpolates it into a prompt, creating a vulnerability surface.
Ingestion points: The code uses requests.get(url) to fetch content from URLs found within the user's message (chat.message).
Boundary markers: While the example wraps the fetched content in <content> tags, it does not provide instructions to the model to ignore any embedded directives or control characters within that content.
Capability inventory: The module has network access via the requests library and the ability to influence LLM completions via llm.stream_final_completion.
Sanitization: The example code does not perform sanitization, filtering, or validation of the remote content before it is processed by the LLM.
[EXTERNAL_DOWNLOADS]: The documentation instructs users to pull and run the ghcr.io/av/harbor-boost Docker image. This is a vendor-provided resource belonging to the skill's author.

boost-modules