The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a mechanism to poll inbound SMS messages via the twilio-inbox command, which introduces a surface for indirect prompt injection from untrusted external senders.\n
Ingestion points: The _twilio_inbox function in scripts/telephony.py retrieves raw message bodies from the Twilio REST API.\n
Boundary markers: The script does not wrap retrieved messages in protective delimiters or provide instructions to the agent to ignore embedded commands within the SMS text.\n
Capability inventory: The telephony script possesses capabilities to send messages, place phone calls, and modify local configuration files (~/.hermes/.env).\n
Sanitization: The retrieved message content is processed and returned to the agent without sanitization or filtering for potential injection patterns.\n- [DATA_EXFILTRATION]: The vapi-import-twilio functionality reads sensitive Twilio credentials (Account SID and Auth Token) from the local ~/.hermes/.env file and transmits them to the Vapi API endpoint (https://api.vapi.ai/phone-number). While this is a documented integration feature, it involves the automated transfer of secrets between different third-party service infrastructures.\n- [COMMAND_EXECUTION]: The skill instructions guide the agent to execute Python scripts where sensitive API keys and authentication tokens are provided as plain-text command-line arguments (e.g., using save-twilio or save-bland). This can result in credential exposure within shell history files or system-level process monitoring tools.

telephony