Lilac Labs — agentic threat model

Not certain from the listing — the specific underlying voice/LLM models are not disclosed, but they are inherently susceptible to voice-based prompt injection, acoustic adversarial perturbations, and mis-aligned verbal outputs to customers.

Not certain from the listing — how customer voice data, menu RAG databases, and order history are stored or processed is undisclosed, risking privacy violations if voice prints are retained or leaked.

The agent directly integrates with existing POS and kitchen systems to send orders, making insecure tool calling, order manipulation, and unauthorized price overrides a primary threat at the framework level.

Not certain from the listing — the deployment architecture (edge hardware at the drive-thru vs. cloud-hosted voice API) is unspecified, leaving potential vulnerabilities in local network security or cloud API endpoints.

Not certain from the listing — while a 98% accuracy rate is claimed, the mechanisms for real-time monitoring, logging of anomalous voice inputs, and output guardrails to prevent brand damage are not detailed.

Not certain from the listing — compliance with PCI-DSS (if handling payments) or voice privacy regulations (GDPR/CCPA/BIPA) is not stated, posing legal and regulatory risks if voice data is captured without explicit consent.

The agent interacts directly with kitchen displays and POS systems, meaning a compromise or failure in the agent can propagate directly to physical kitchen operations, causing denial of service or fraudulent transactions.