The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill retrieves external content from SMS messages and call transcripts that may contain adversarial instructions, creating a surface for indirect prompt injection. \n
Ingestion points: Incoming SMS messages are fetched via the _twilio_inbox function in scripts/telephony.py. Call transcripts are retrieved via _bland_status and _vapi_status in scripts/telephony.py after AI-driven calls are completed. \n
Boundary markers: The script does not wrap retrieved content in protective delimiters or provide explicit instructions for the agent to ignore embedded commands within the fetched data. \n
Capability inventory: The skill possesses significant capabilities, including sending SMS/MMS, initiating automated outbound calls, and purchasing phone numbers. \n
Sanitization: No sanitization, filtering, or validation is performed on the retrieved message bodies or transcripts before they are provided to the agent context. \n- [DATA_EXFILTRATION]: The skill facilitates the transmission of sensitive credentials to external service providers as part of its integration features. \n
Evidence: The vapi-import-twilio command in scripts/telephony.py reads the user's Twilio Account SID and Auth Token and transmits them to the Vapi API (api.vapi.ai) to enable number integration. \n- [COMMAND_EXECUTION]: The skill relies on the execution of a local Python helper script that interacts with the user's environment and filesystem. \n
Evidence: The skill instructions prompt the agent to execute scripts/telephony.py using the Python interpreter to manage credentials and telephony state in the ${HERMES_HOME}/.env and telephony_state.json files.

telephony