The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user or agent to download and install a CLI tool from a third-party GitHub repository (Virtual-Protocol/openclaw-acp).
[COMMAND_EXECUTION]: The workflow relies on executing shell commands such as acp setup, acp browse, and acp job create using an external tool.
[COMMAND_EXECUTION]: User-provided data (cpdv_data) is interpolated directly into a shell command string. This pattern is vulnerable to command injection if the input is not correctly escaped, despite the documentation providing a manual instruction to do so.
[DATA_EXFILTRATION]: The skill transmits private data to a remote provider wallet to facilitate Zero-Knowledge Proof generation. While this is the intended purpose, it involves sending sensitive information to an external system.
[PROMPT_INJECTION]: The skill lacks sufficient safeguards for processing untrusted external data in a high-privilege command context.
Ingestion points: cpdv_data parameter in SKILL.md and evals.json.
Boundary markers: Absent; user input is placed directly within the command arguments.
Capability inventory: Shell command execution via the acp CLI tool.
Sanitization: No programmatic sanitization or validation is implemented; it relies on manual user or agent compliance with escaping instructions.

private-data-verification