The Agent Skills Directory

[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by instructing the agent to ingest and process content from external sources.
Ingestion points: The agent is instructed to search and read 'live' author instructions from external journal websites and process user-provided manuscripts.
Boundary markers: No specific delimiters or instructions are provided to the agent to distinguish between its own system instructions and potentially adversarial content within the external data.
Capability inventory: The skill requires the agent to use web search and file reading capabilities.
Sanitization: There are no requirements for the agent to validate, filter, or sanitize the external data retrieved during the workflow.
[DATA_EXFILTRATION]: The skill uses relative paths to access resources outside its own directory, which is a pattern associated with data exposure.
Evidence: In the 'Official-submission checklist' section, the agent is instructed to read ../../resources/source-basis.md and ../../resources/official-source-map.md. Accessing files via relative path traversal (../../) can be used to probe or access files outside the intended execution scope.

the-astrophysical-journal