The Agent Skills Directory

[Data Exfiltration] (HIGH): The skill provides the capability to read local files via the --file argument and send their contents to any URL. An attacker could use indirect prompt injection to trick an agent into exfiltrating sensitive files (e.g., credentials or SSH keys) to a malicious server.
[Indirect Prompt Injection] (HIGH): Mandatory Evidence Chain:
Ingestion points: Processes data from arbitrary external URLs via the requests library in scripts/http_client.py.
Boundary markers: None present; the tool treats all response data as trusted for display or storage.
Capability inventory: Arbitrary network requests (GET/POST/etc.), file reading (--file), and file writing (--output).
Sanitization: None; the tool provides raw access to the network and filesystem without validation of URLs or file paths.
[Command Execution] (MEDIUM): While the script itself does not call eval() or subprocess, the AI agent's use of this script involves executing a CLI tool that can modify the local filesystem (--output), which could be leveraged for persistence or overwriting critical configuration files.

http-client