The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a local bash script scripts/crawlora.sh to interface with the Crawlora API. The script is highly defensive, incorporating input validation via anchored regular expressions to whitelist specific API paths and methods. It also correctly sanitizes inputs to prevent local file disclosure by blocking the @ character in query arguments and using stdin for POST bodies to avoid curl interpreting input as a filename. Credential handling is secured by writing the API key to a temporary file with restricted permissions, passed to curl via a configuration file to avoid exposure in process lists.
[INDIRECT_PROMPT_INJECTION]: The skill processes external content retrieved from websites and research papers, which presents an attack surface for indirect prompt injection.
Ingestion points: Data fetched via scripts/crawlora.sh from the /web/search, /web/crawl, and /research/papers/read endpoints in SKILL.md.
Boundary markers: The skill instructions lack specific delimiters or "ignore" directives for isolating the processed external content, though it does emphasize citing exact quotes.
Capability inventory: The skill possesses network communication capabilities with api.crawlora.net and shell execution of the bundled helper script.
Sanitization: While the wrapper script validates API request parameters (path and methods), it does not sanitize the returned text content of search results or papers for embedded instructions targeting the LLM.

cross-corpus-research