The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the openclaw CLI tool for initiating and managing voice calls. This involves executing shell commands which, if misconfigured, could lead to unintended system interactions.\n- [DATA_EXFILTRATION]: Handles sensitive metadata including phone numbers and call IDs, as well as credentials for third-party providers (Twilio, Telnyx, Plivo). The skill contains specific guardrails instructing the agent never to pipe this output to external commands or network endpoints, mitigating the risk of metadata harvesting.\n- [PROMPT_INJECTION]: Vulnerable to indirect prompt injection where an attacker might provide a malicious message string to be spoken by the agent. This is addressed by a mandatory requirement for the user to author or explicitly approve any message relayed via the tool.\n
Ingestion points: Untrusted data enters the agent context through the message argument in the initiate_call, continue_call, and speak_to_user tool actions within SKILL.md.\n
Boundary markers: Security instructions are explicitly defined in the Security Guardrails section of SKILL.md to prevent accidental execution of injected instructions.\n
Capability inventory: Subprocess calls via openclaw CLI for telephony operations and status checks across all provided tool definitions.\n
Sanitization: The skill relies on human-in-the-loop verification and explicit user approval for all outbound communication rather than automated programmatic sanitization.

voice-call-hardened