Voxworks — agentic threat model

Not certain from the listing — Voxworks likely utilizes third-party LLMs and text-to-speech/speech-to-text models. Threats include voice-based prompt injection (audio injection/vishing), model reprogramming, and adversarial audio inputs designed to bypass system prompts.

Voxworks guarantees local Australian data residency. Threats include unauthorized access to stored call recordings, transcripts, or customer PII within the local hosting environment, as well as potential data exfiltration if the local storage is misconfigured.

Not certain from the listing — The agent framework orchestrates voice call flows and likely integrates with external business tools (CRMs, calendars). Threats include insecure tool integration, state manipulation during live calls, and memory poisoning if call history is used to influence future interactions.

Not certain from the listing — The platform is hosted locally in Australia. Threats include SIP trunk hijacking, telephony infrastructure compromise, and unauthorized access to the hosting servers or API endpoints connecting the voice agent to the telecom network.

Not certain from the listing — It is unclear what real-time guardrails or monitoring are in place. Threats include a lack of low-latency voice guardrails, allowing users to socially engineer the agent or exploit conversational drift without immediate detection.

Voxworks emphasizes data sovereignty and local Australian residency, aligning with Australian Privacy Principles (APPs). Threats include compliance violations if downstream integrations or sub-processors transfer data outside Australia, or if access controls to call logs are insufficient.

Not certain from the listing — The agent operates primarily as a standalone telephony solution. Threats are limited to horizontal integrations (e.g., CRM, booking systems) where a compromise of the voice agent could lead to cascading unauthorized actions in connected business systems.