The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the Bluesky social network (profiles, posts, feeds) via the Crawlora API.
Ingestion points: Data enters the agent's context through scripts/crawlora.sh which returns JSON-formatted content from external API calls.
Boundary markers: The instructions specify that the tool returns "clean JSON" and "normalized JSON," which helps the agent distinguish between data and instructions, though explicit "ignore embedded instructions" markers are not present in the prompt template.
Capability inventory: The skill possesses network access to the vendor's API (api.crawlora.net) and file-read access to its own helper scripts. It does not have file-write capabilities or arbitrary shell execution.
Sanitization: The scripts/crawlora.sh helper uses curl --data-urlencode for all query parameters, preventing basic command or parameter injection. It also strictly validates the API key format and restricts requested paths to a whitelist of allowed routes.
[SAFE]: The skill follows security best practices for credential management and input validation.
Credential Safety: The skill uses an environment variable (CRAWLORA_API_KEY) and writes it to a temporary, restricted-permission curl configuration file (chmod 600) to avoid leaking the key in process lists or shell history.
Resource Validation: The helper script implements a strict path whitelist and rejects malformed paths (e.g., those containing glob characters or directory traversal patterns) before execution.

bluesky-research