The Agent Skills Directory

[PROMPT_INJECTION]: The skill is inherently susceptible to indirect prompt injection because it is designed to ingest and process untrusted data from the open web.
Ingestion points: External content from various websites is retrieved via the search, research, and extract commands in scripts/parallel_web.py and passed back to the agent context.
Boundary markers: There are no explicit delimiters or specific instructions (e.g., XML tags or 'ignore embedded instructions' markers) wrapping the retrieved web content when it is displayed or saved.
Capability inventory: The skill is granted Write and Bash permissions, which allow it to create and modify files in the local sources/ directory based on the research results.
Sanitization: The implementation in scripts/parallel_web.py does not include validation or sanitization layers to filter potential malicious instructions embedded in the HTML or text returned by the Parallel API.

parallel-web