The Agent Skills Directory

[DATA_EXFILTRATION]: The request.sh script provides the ability to include local file contents in the request body or as form data using the --file and --form (with @ syntax) flags. This creates a vector for exfiltrating sensitive local information, such as configuration files, environment variables, or private keys, if the agent is directed to process these paths.
[COMMAND_EXECUTION]: While the skill uses shell arrays to safely pass arguments to curl, the tool itself provides a powerful network communication capability that can be used to interact with arbitrary endpoints from the local host environment.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from external HTTP responses.
Ingestion points: HTTP response bodies returned to stdout or saved to files via the --out option.
Boundary markers: No delimiters are used to separate external response data from the agent's internal instructions.
Capability inventory: The agent can use this skill to perform further network requests, read/write files, and execute other tools based on the content of the response.
Sanitization: The skill does not sanitize or validate response content before returning it to the agent.
[EXTERNAL_DOWNLOADS]: The index.js implementation includes hostname validation to block private and internal IP ranges (SSRF protection). However, the request.sh shell script does not implement any such validation, allowing requests to be directed toward internal network resources or cloud metadata services (e.g., 169.254.169.254).

http-request