Call Support AI — agentic threat model

Not certain from the listing — likely relies on third-party LLMs combined with Speech-to-Text (STT) and Text-to-Speech (TTS) engines. It is highly vulnerable to voice-based prompt injection (vishing injection) where callers manipulate the underlying LLM to bypass screening or extract system prompts.

Not certain from the listing — processes real-time voice inputs, customer feedback, and claims data. Without explicit details on data retention, there are risks of PII leakage, unauthorized data exfiltration, and lack of encryption for stored call transcripts or audio recordings.

Not certain from the listing — orchestrates conversational state, schedules viewings, and forwards information. Insecure tool integration with CRMs or scheduling APIs could allow an attacker to manipulate database records or trigger unauthorized actions via conversational exploits.

Not certain from the listing — requires telephony infrastructure (SIP/WebRTC gateways) and hosting. Vulnerabilities include SIP toll fraud, denial of service on voice lines, and insecure storage of API keys used to connect to telephony providers.

Not certain from the listing — lacks details on real-time guardrails or call monitoring. Without robust observability, malicious interactions, prompt injections, or model drift during live calls may go undetected.

Not certain from the listing — handles sensitive verticals like real estate and insurance claims but mentions no compliance frameworks (e.g., SOC2, GDPR, HIPAA). This poses significant regulatory and privacy compliance risks regarding caller consent and data handling.

Not certain from the listing — interacts with downstream systems and 'relevant people' by forwarding information. If integrated into a multi-agent workflow or automated dispatch system, compromised inputs could cause cascading logic failures in connected business systems.