The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from scratch/audit/<stem>/audit.md and content retrieved from documentation_url to generate experiment backlog items. This creates a surface for indirect prompt injection if the source data or fetched web pages are compromised.
Ingestion points: The local file scratch/audit/<stem>/audit.md and external web content fetched via WebFetch from the documentation_url field.
Boundary markers: There are no explicit instructions or delimiters defined to prevent the agent from following malicious instructions potentially embedded in the fetched documentation content.
Capability inventory: File read operations, network fetch (WebFetch), and the ability to trigger the audit-ml-pipeline runner.
Sanitization: No validation or sanitization of the remote content is performed before interpolating it into the agent's context to draft backlog items.
[EXTERNAL_DOWNLOADS]: The skill follows the documentation_url links found in the audit report using the WebFetch tool. In the provided documentation, these URLs point to the vendor's domain (docs.skore.probabl.ai), but the skill logic itself does not restrict requests to a specific domain, allowing it to fetch content from arbitrary external sources.

iterate-from-skore